Sophie Well

Deterministic healthcare utilities, free forever, no servers, no accounts.

sophiewell.com is a single-page static website built for the nurse on shift — primarily ICU and acute-care RN (CCRN / PCCN), then ED RN (CEN), then floor / med-surg RN, then OR / PACU RN, then L&D / NICU RN — with every relevant tile still served to doctors, pharmacists, RTs, billers, coders, and EMS providers. It is, in posture, MDCalc with no ads, no login, no upsell, no cookie banner, no email capture, and no network call after first paint. Everything runs in the browser. There is no AI of any kind. The data sources are public, bundled with the page, and refreshed on a schedule. The site costs nothing to operate beyond the domain renewal, so it can be free forever. The product thesis, in one sentence: MDCalc, but free, login-less, ad-less, offline-capable, with a catalog tuned to the nurse on shift (see docs/spec-v29.md). The one-line scope test: a tile must consume at least one user input and produce a computed output; "searchable lookup of static facts" does not qualify. See docs/spec-v10.md for the audience and dependency-budget commitments and docs/spec-v29.md for the nurse-first pivot and the v29 catalog ledger. At v277 close the catalog is 1137 deterministic tiles — every one of them computes from at least one user input. The catalog reached its present size on two tracks. New tiles: spec-v63 added the operations counterpart to the bedside surface — a shared regulatory-deadline engine (lib/deadline.js) and five calculators (Medicare appeal-level deadlines, claim timely-filing, the 2021 E/M Medical-Decision-Making level, the prior-authorization decision clock, and the 60-day overpayment clock); spec-v64 added calcium-replacement, the IV-calcium / elemental-calcium / gluconate↔chloride converter that closes the one electrolyte the K/Mg/Phos electrolyte-replacement ladder omitted; and spec-v65 added three bedside-physiology calculations a nurse still does on scratch paper (o2-cylinder-duration, minute-ventilation, and cerebral-perfusion-pressure, CPP = MAP − ICP). Zero-new-tile hardening (spec-v63 Part A, spec-v66 → spec-v76): correctness fixes that aligned each printed band to what the code actually computes — abg Boston-rules compensation, the symmetric over-rapid-correction warning in acid-base-deficit, the ttkg renal-wasting threshold, the indication-aware digoxin floor, the sas-riker light-sedation goal band, and the reachable psi Risk Class I — alongside denial→next-step ops routing, CFR-checklist document generators, and a run of accessibility / rendering / offline repairs (44px touch targets, theme-tracked color-scheme, a working SPA skip-link, a complete-shell service-worker precache, and a guarded tool-page discovery allowlist). Per-wave detail lives in the CHANGELOG and the spec docs (spec-v62 through spec-v76, with the offline shell precache finished in spec-v84 — every local asset index.html loads is now precached, guarded so the list cannot drift). Advanced clinical calculators (spec-v85 program): the spec-v85 charter opens a fourteen-spec program that deepens the physician-facing critical-care / subspecialty layer one rung above the bedside score, under a binding doctrine (one-line determinism, no new bundled dataset, primary-source citations, the output-safety contract) and a CI/CD maintenance contract (build pipeline, merge gates, Class A/B staleness machinery). Its first feature spec, spec-v86, ships three deterministic toxicology decision rules — serotonin-toxicity (Hunter criteria), salicylate-toxicity (EXTRIP hemodialysis indication), and toxic-alcohol (ethanol-corrected osmolar gap + AACT fomepizole indication). spec-v87 follows with three critical-care physiology calculators in Group E — hemodynamic-suite (the PA-catheter cardiac-index / SVR / PVR resistance suite, PVR reported in both dynes·s·cm⁻⁵ and Wood units per ESC/ERS 2022), mechanical-power (the Gattinoni simplified power of ventilation with the >17 J/min VILI-risk flag), and dead-space (the Bohr-Enghoff Vd/Vt fraction with the EtCO₂-surrogate caveat). spec-v88 adds three high-acuity endocrine/oncology calculators — dka-hhs (the ADA hyperglycemic-crisis classification: DKA vs HHS with mild/moderate/severe DKA grading, plus the computed anion gap and effective serum osmolality, Group G), calvert-carboplatin (the AUC-based carboplatin dose by the Calvert formula with the FDA estimated-GFR cap at 125 mL/min shown as a visible substitution, Group F), and tls-cairo-bishop (the Cairo-Bishop tumor-lysis-syndrome laboratory/clinical grading with the 25%-change-from-baseline branch and the corrected-calcium criterion, Group G). spec-v89 closes Wave 1 of the spec-v85 program with four subspecialty calculators (all Group G) — das28 (the DAS28-ESR/DAS28-CRP rheumatoid-arthritis disease-activity score with the EULAR remission/low/moderate/high bands, the catalog's first rheumatology tile), kings-college (the King's College Criteria for transplant referral in acetaminophen-induced acute liver failure, with the pH limb, the three-part coagulopathy/renal/encephalopathy limb, and the Bernal lactate modification), asa-ps (the ASA Physical Status classification I–VI with the E-modifier rules enforced), and surgical-apgar (the Gawande intraoperative 0–10 outcome score, distinct from the neonatal Apgar). spec-v90 opens Wave 2 with six cardiology / ECG computations — ecg-axis (the mean frontal-plane QRS axis by hexaxial atan2 geometry, with the all-isoelectric (0,0) input surfaced as "indeterminate axis" rather than a spurious 0°, Group E), lvh-criteria (the Sokolow-Lyon and Cornell ECG-LVH voltage criteria with the sex-specific Cornell threshold, Group G), timi-stemi (the Morrow 2000 TIMI risk score for STEMI with the 30-day mortality band, Group G), duke-treadmill (the Mark 1987 exercise-test prognosis with the cited five-year survival, Group E), cardiac-power-output (the Fincke CPO = MAP×CO/451 watts with the <0.6 W cardiogenic-shock threshold, Group E), and aortic-valve-area (the continuity-equation valve area with the dimensionless index and the ASE/EACVI 2017 + 2020 ACC/AHA severity bands, Group E). As the first Wave-2 spec it also authors the spec-v85 §6.3 warn-only monthly scripts/check-citation-cadence.mjs job, which annotates (but never blocks) when a calendar-tracked Class-B citation row is overdue for re-verification. spec-v91 continues Wave 2 with five pulmonary-function / chronic-respiratory computations that fill the gap beside the catalog's acute respiratory surface (aa-pf-suite, rox, curb-65, smart-cop) — gold-spirometry (the GOLD spirometric COPD grade off post-bronchodilator FEV1/FVC < 0.70 and FEV1 %predicted, Group G), bode-index (the Celli 2004 multidimensional COPD prognosis 0–10 with the 4-year survival quartile, Group G), gap-ipf (the Ley 2012 GAP index for idiopathic pulmonary fibrosis with the cannot-perform-DLCO limb and stage mortality, Group G), predicted-spirometry (the GLI-2012 LMS predicted FEV1/FVC/ratio + lower limit of normal from compiled coefficient/spline constants, Group E), and mmrc-dyspnea (the Bestall 1999 modified MRC dyspnea grade 0–4 that feeds BODE and the GOLD ABE assessment, Group G). spec-v92 continues Wave 2 with five nephrology computations that close the chronic / procedural renal gap beside the catalog's existing filtration / injury / dosing surface (egfr-suite, fena-feurea, kdigo-aki, cockcroft-gault) — ckd-staging (the KDIGO 2024 CKD G×A risk heat-map cell from eGFR and UACR, Group G), uacr-upcr (the spot urine albumin/protein-to- creatinine ratios with the estimated 24-hour excretion and the KDIGO A-stage, Group E), ktv-urr (the hemodialysis adequacy URR + Daugirdas second-generation single-pool Kt/V against the KDOQI targets, Group E), mehran-cin (the Mehran 2004 contrast-induced-nephropathy risk score with the CIN / dialysis bands, Group G), and ckd-epi-cystatin (the 2021 race-free CKD-EPI cystatin-C / combined / creatinine eGFR, Group E). The new pa-lint tile in spec-v52 consumes dropped files instead of form fields and produces a deterministic findings report, the first instance of the shape: 'document-linter' tile shape. Catalog-truth invariants (docs/spec-v46.md) fail CI on any drift between UTILITIES.length and the public marketing copy. Sophie's eight posture commitments (docs/spec-v50.md) — no ads, no login, no telemetry, no third-party fetch, no AI, no cookies, no paid tier, MIT-licensed forever — are listed at /commitments/ and enforced by automated checks on every commit. Scoring tiles expose a collapsed "where does this come from?" derivation block (docs/spec-v48.md) — 126 carry one today — with the verbatim source formula, study cohort, limits of validity, and a live per-input contribution list whose component sums are cross-checked in CI against the tile's own scoring function.

The problem

Bedside math — drug dose, drip rate, anion gap, eGFR adjustment, sepsis-bundle clock, restraint re-check, Braden re-score — lives in published clinical literature and institutional protocols. The nurse on shift, the resident at 3 a.m., the medic in the ambulance, and the pharmacist verifying a renal dose all reach for the same calculators and arrive at the same number from the same inputs. Existing options are paywalled, ad-laden, login- gated, account-tied, or quietly telemetered — every one of which adds friction to a 30-second decision at the bedside.

The meta-problem is that the workers who would benefit most from fast, free, deterministic math are the ones least likely to have a paid app handy at the moment they need it.

Related MCP server: Chia Health MCP

The solution

Take the public datasets and the published clinical formulas, ship them to the browser, do the lookups and math locally, and never phone home. A single static page, a Content Security Policy that forbids outbound network connections, and a service worker that caches the page for offline use.

Quick start

Visit https://sophiewell.com. Save the page for offline use if desired. To run locally, clone the repository and run npm run dev, which starts a zero-dependency static server on http://localhost:4173 that applies the production security headers. Any static file server will also work.

How it works and how to use it

Since the spec-v29 nurse-first prune the catalog has grown one reviewable spec at a time to 1137 deterministic calculators (the full per-version history is in CHANGELOG.md and docs/spec-v*.md; the most recent bedside additions are summarized in the cheat sheets below). They organize across the bedside-shift surfaces a nurse, doctor, pharmacist, RT, EMS provider, biller-coder, or educator actually reaches for. Every tile takes at least one user input and produces a computed output; searchable indexes of static facts are explicitly out of scope (see docs/spec-v29.md §3 and docs/spec-v10.md §2.3). Tiles that sit in the same clinical workflow cross-link: a "Related tools" row — present across most of the catalog (spec-v61 §2 A2) — puts the sibling a nurse reaches for next: wells-pe → perc / pesi / years-pe, cockcroft-gault → the renal-dosing tools, one tap away, with every link verified in CI to resolve to a real tile.

Clinical math & conversions covers BMI, BSA suite, MAP / pulse pressure / shock index, anion gap with delta-delta, corrected Ca / Na, osmolal gap, A-a gradient and P/F suite, Winter's formula, eGFR suite (CKD-EPI 2021 / MDRD / Cockcroft- Gault), FENa / FEUrea, maintenance fluids (4-2-1), QTc suite, pregnancy dating, pack-years, the universal unit converter, sodium-correction planner (Adrogue-Madias), free water deficit, predicted body weight + ARDSnet tidal volume, and RSBI. Medication & infusion covers drip rate, weight-based dose, concentration-to-rate, pediatric dose bounds, insulin drip, anticoagulant reversal, opioid MME (CDC 2022), steroid and benzodiazepine equivalence, antibiotic renal-dose adjustment, vasopressor dose-to-rate (with VIS / Wernovsky IS), TPN macronutrient, iron deficit (Ganzoni), and the v29 nursing- shift additions for insulin correction, electrolyte replacement ladders, CRRT dose, ECMO titration, and the MTP ratio tracker. Clinical scoring & risk covers GCS, APGAR, NIHSS / mNIHSS, Wells PE / DVT, CHA2DS2-VASc, HAS-BLED, TIMI, GRACE, HEART, PERC, Geneva, CURB-65, PSI, qSOFA / SOFA, MELD-3.0 / Child- Pugh, Ranson / BISAP, Centor / McIsaac, Caprini, Bishop, Alvarado / PAS, PHQ-9, GAD-7, AUDIT-C, CAGE, EPDS, Mini-Cog, CIWA-Ar, COWS, ASCVD PCE, PREVENT 2023, Light's criteria, Mentzer index, SAAG, R-factor liver injury, KDIGO AKI, ICH Score, Hunt-Hess / WFNS, plus modified Sgarbossa, revised cardiac risk index, PEWS / NEWS2 / NEWS2-escalation, every v17-v28 risk-score tile, and the v29 bedside nursing screens (Braden, Morse, Hendrich II, RASS, BPS, CPOT, ICDSC, CAM, CAM-ICU, Aldrete / PADSS). Clinical criteria & diagnostic bundles packages NPIAP pressure-injury staging, Norton / PUSH wound assessment, the VIP / INS extravasation criteria, ABO / Rh blood-product compatibility, and the v17-v28 diagnostic bundles. Workflow & templates carries the patient-visit generators that survive v29 (appeal letter, HIPAA Right of Access, HIPAA authorization, ROI request, discharge instructions, specialty- visit questions, the wallet-card generator, and the SBAR handoff template), plus the v29 nursing-shift workflow timers: restraint timer, sepsis-bundle clock, code-blue clock, device- day counter, Bristol / abdominal-girth tracker, and the vent SBT readiness / ARDSnet PEEP-FiO2 ladder. Field medicine covers NEXUS / Canadian C-Spine, CDC Field Triage, START / JumpSTART, peds-weight-dose, burn surface area and fluid resuscitation (Parkland / modified Brooke), pediatric ETT sizing, naloxone dosing, the EMS documentation helper, and the AVPU / GCS quick reference. Public health & infectious disease covers tetanus prophylaxis, rabies PEP, bloodborne pathogen exposure, TB testing interpretation, and STI screening intervals. Billing & coding is now sparse — the time-based E/M code selector, the NDC 10 ↔ 11 digit converter, and the HIPAA 60-day breach clock. Every static index (ICD-10-CM, HCPCS, CPT, NDC, POS, modifier, revenue, CARC / RARC, NUBC, DRG, APC, ICD-10-PCS, RxNorm, NDC↔RxNorm) was retired in spec-v29 wave 29-2 §2.1; use your EHR, CMS, FDA, or NUBC source instead.

The user flow is simple: type what you need into the hero search ("wells PE", "CHA2DS2-VASc", "ICD-10", "magnesium replacement") or follow one of the static browse links below it, enter input, read output. The hero consults a hand-curated synonym table (data/synonyms.json, spec-v7 section 3.2) before falling back to fuzzy matching, and shows a one-line breadcrumb explaining why a tile is recommended. (The spec-v7 §3.1 patient-artifact dropzone was retired in the 2026-05-18 clinical-staff-first pivot; the deterministic classifier under lib/artifact-*.js is retained for possible future reuse on a clinical-input surface, but the home view no longer wears it.) Every utility opens with its inputs pre-filled from a worked example so the empty state is never empty (a "Reset to example" link restores them after editing), and the References region at the bottom shows the formula citation or dataset stamp the result came from (spec-v9). Calculator state is encoded in the URL hash so any view can be bookmarked or shared as a permalink.

All computation happens in the browser. For the full picture, see docs/architecture.md.

Tile anatomy: one calculator, four single-sources-of-truth

Every tile is assembled from four files, each owning exactly one concern, so a fact is typed once and read everywhere. Nothing is duplicated across the runtime, the SEO build, or the optional MCP surface — the catalog-truth and MCP-catalog gates fail the build if any of them drift.

                 ┌──────────────────────────────────────────────┐
   one tile id ──┤  app.js  UTILITIES[]   name · group · clinical│  what & where
                 └──────────────────────────────────────────────┘
                       │ id                         │ id
                       ▼                            ▼
   ┌───────────────────────────────┐   ┌──────────────────────────────────┐
   │ lib/meta.js  META[id]         │   │ views/group-*.js  RENDERERS[id]   │
   │  citation · citationUrl       │   │  builds the DOM inputs + wires the │
   │  worked example · bands       │   │  live recompute (pure el(), no     │
   │  specialties · disclaimer     │   │  innerHTML)                        │
   └───────────────────────────────┘   └──────────────────────────────────┘
                                                    │  reads inputs via
                                                    ▼
        unitField(label,id,UNITS) ─► unitNum(id) / unitNumOpt(id)
              (°C|°F · cm|in · kg|lb · mg/dL|mmol/L; canonical = default option,
               so the documented example reproduces byte-identically)
                                                    │ canonical-unit value
                                                    ▼
                       ┌────────────────────────────────────┐
                       │ lib/*.js  pure compute fn           │
                       │  routes every divide through        │
                       │  lib/num.js → finite, never NaN     │
                       └────────────────────────────────────┘
                                                    │ result object
                                                    ▼
                          resultRow() → bands → references (citation/stamp)
                                                    │
        ┌───────────────────────────────────────────┴───────────────────────┐
        ▼                                                                     ▼
  state encoded in the URL hash (bookmark / share / deep-link)     fuzz harness +
  static SEO page (build-tool-pages) + optional MCP tool           example-correctness
  (mcp/adapters/*.js) all read the SAME compute + META             e2e assert the same
                                                                    numeric contract

The unit-toggle row (unitField → unitNum) is the seam that lets a US nurse chart in °F / inches / lb while the compute function still receives canonical metric: the canonical unit is always the first <select> option, so every META.example and every shared hash reproduces a calculation byte-for-byte.

Bedside-math cheat sheet (spec-v55 additions)

The thirteen Group-E calculators added in spec-v55, with the formula a nurse or RT would otherwise run on scratch paper. Every denominator is guarded (a bad input shows a (…) fallback, never a non-finite number), and each ships its primary citation inline on the tile.

Dosing & infusion cheat sheet (spec-v56 additions)

The thirteen Group-F medication/infusion calculators added in spec-v56. Each is dosing decision-support, not a prescription, and renders the standing "verify against institutional protocol and a current reference" notice. Two tiles refuse outside their validity window rather than mislead.

Screeners & decision-rule cheat sheet (spec-v57 additions)

The fourteen Group-G instruments added in spec-v57. Each is screening / decision support, not a diagnosis. Two surface a conditional threshold so the user sees why the determination flipped.

Neonatal, maternal & ICU cheat sheet (spec-v58 additions)

The twelve neonatal / maternal / pediatric-and-adult-ICU scores added in

spec-v58, closing the 50-tile expansion begun in v55 (255 → 307). Each is a published instrument a bedside nurse already assigns by hand. Three apply age- or gestational-age-banded cutoffs automatically and show the active band; three neonatal scores and braden-q state their direction (higher-vs-lower = worse) explicitly so a cross-reading nurse cannot invert it.

Medication-safety, electrolyte/fluid & OB/peds cheat sheet (spec-v61 additions)

Twelve bedside computations a nurse otherwise does by hand — the v61 wave (307 → 319). Each computes an output from input (passes the spec-v29 §3 one-line test), ships its primary citation inline with a DOI, and renders an explicit "estimate / verify per local protocol and an independent double-check" note on every dosing/replacement tile: the order stays with the clinician and the pharmacy.

The wave also added related-tool links (META[id].related, rendered as a "Related tools" row in the citation block — e.g. wells-pe → perc / pesi / years-pe). The rollout is now complete: a single reviewable RELATED_BACKFILL map (spec-v61 A2) carries the linking across the catalog — 314 curated sibling clusters, grouped by clinical family — with every link verified in CI to resolve to a real tile and each list capped at four siblings so the row never crowds the result at 320px. The wave also added a "Copy link" affordance next to "Copy all" that copies the deep link (hash-state already encodes the inputs), so a populated calculation can be handed to a colleague with no new persistence and no network. The shared printable template (lib/print.js, with its "No data was sent or stored" footer) now also covers the SBAR handoff and code-blue summary tiles (spec-v61 A6), so a nurse can print a structured handoff or a code timeline straight from the tile.

Interpretation-band parity (spec-v61 A8). Every backfilled score now shows a source-anchored "Per source:" band block under its citation — the verbatim meaning of the number from the instrument's own paper, not Sophie's phrasing. The pass raised META[id].interpretation coverage from 150 to 196 of the catalog's scores, covering the recent bedside scores plus classics (APGAR, qSOFA, MELD, Ranson, Alvarado, AUDIT-C, ASCVD/PREVENT, KDIGO-AKI, ARISCAT, APACHE II, Braden Q, and more). The bands are authored as one reviewable merge map in lib/meta.js and render through the shared renderMetaBlock with zero per-view wiring; a CI guard (test/unit/meta-interpretation.test.js) pins every band to sourceQuoted: true, a non-empty sourceCitation, ≤200 chars, and no Sophie-authored phrasing. A second invariant in the same guard requires every tile that carries discrete derivation.bands (a score with named result cut-points) to also carry an interpretation block, so the "where the number comes from" and "what it means per source" displays cannot drift apart; continuous-mortality scores (pelod2, psofa) omit discrete bands and stay out of the rule. The final tile to satisfy this was pews (Brighton PEWS), whose interpretation restates its Monaghan-2005 escalation thresholds.

Opt-in input persistence (spec-v61 A7). Tiles with numeric/choice inputs show a "Remember my inputs on this device" toggle in the references block, off by default. When a nurse opts in, that tile's values are written to localStorage (lib/input-persist.js) so reopening it next shift skips re-entering constants. Only number/range/checkbox/radio inputs and <select> values are stored — free-text and <textarea> are never persisted, so a name, allergy, or clinical note cannot reach storage. The two keys (sw-remember, sw-saved-inputs) are string literals on the storage allowlist enforced by check-commitments; unchecking the toggle erases both. Remembered values fill fields a deep link did not set and win over the example. Nothing leaves the device — the CSP still blocks every network egress.

Unit toggles & chart-ready copy (spec-v61 A4/A3). The Group E clinical-math tiles now meet the nurse in her own units. bmi, bsa, and cockcroft-gault carry a per-field unit <select> (weight kg⇄lb, height m/cm⇄in, creatinine mg/dL⇄µmol/L) driven by the existing lib/unit-convert.js converters; each option converts to the canonical unit before the formula runs, so the math is untouched and — because the canonical unit is always the default — every documented example and deep-link hash reproduces a calculation byte-identically. The input+select row wraps, so it never forces horizontal scrolling on the narrowest phones. Alongside, the multi-output tiles build their results as {label, value, units} items and render a "Copy results" button that pastes clean Label: Value Units lines via lib/clipboard.js formatCopyAll — a chart-ready paste instead of a scraped innerText blob. The labeled copy covers the Group E math tiles (bsa, anion-gap, corrected-sodium, aa-gradient), the five 2+-numeric-output v61 bedside tiles (ebv-mabl, peds-transfusion-volume, rhig-dose, fluid-balance, carb-insulin-bolus), the six multi-output Group V5 lab / clinical-math tiles (sodium-correction, free-water-deficit, iron-ganzoni, pbw-ardsnet, lights, corrected-anion-gap), the three Group F medication / infusion tiles (drip-rate, tpn-macro, insulin-correction, the latter two folding their headline total into the copied list as anion-gap-dd does), and — through the same shared resultRow helper — the three Group I field-medicine tiles (burn-fluid, peds-ett, naloxone, where burn-fluid folds the Parkland and Modified-Brooke schedules into one copyable block) and the five Group V7 oxygenation / renal-acid / lipid tiles (ldl-calc, cao2-do2, oxygenation-index, driving-pressure, acid-base-deficit, each emitting two-to-four computed numeric results) — exactly the values a nurse or medic pastes into a transfusion, I&O, electrolyte-correction, ventilator, pump, TPN, insulin, burn-resuscitation, airway, lipid, or oxygen-delivery chart, with on-screen text byte-identical to the prior hand-built list. All twenty-six are pinned by test/integration/unit-toggle.spec.js (alternate-unit parity, example-prefill parity, the labeled-copy affordance, and a 320px no-overflow assertion).

The A4 weight toggle then rolled out to the dosing tiles. The helpers were extracted to lib/field-units.js and a kg⇄lb toggle added to every weight-bearing dosing tile in Group F (weight-dose, conc-rate, vasopressor, crrt-dose, ecmo-titration) and the twelve v61 bedside tiles (urine-output, gir, ebv-mabl, potassium-deficit, peds-transfusion-volume, burn-uop-target, fluid-balance) — so a US nurse who weighs a patient in pounds skips the hand-conversion before a weight-based dose, infusion rate, or hourly urine-output check. The converter feeds the canonical kg value to both the formula and the boundsAdvisory() plausibility note, so result and advisory match the kg entry exactly; the cross-group lb-parity test in unit-toggle.spec.js pins it. The A1 derivation tail is being worked down one verified wave at a time (the additive screeners sirs, apfel, and aims65; then the ED scores feverpain, canadian-syncope, and stone-score; then padua, epworth, and nrs2002; then the 0-2-per-sign bedside scores apgar, silverman-andersen, and downes; then the PE-prognosis scores pesi and spesi with the pediatric nigrovic meningitis score; then the GI-bleed risk family gbs, rockall, and oakland (the first banded-weight wave, each band encoded as a points callback); then the ICU-prognosis additive indices nutric, mnutric, and mods; then the specialty point scales burch-wartofsky, ariscat, and braden-q; then the high-value scores hacor, vis, and charlson; then the array-scored screeners audit-full, dast10, and gds15 (their reverse-scored items made explicit); then the pediatric/neonatal bedside scales nips, cries, and peds-gcs; then the age-banded pediatric organ-dysfunction scores pelod2 and psofa; then apache2 (the APACHE II adult-ICU severity score, twelve banded variables); then mnihss (the modified NIHSS stroke scale); then finnegan (the modified Finnegan neonatal-abstinence score) — 112 additive scores now carry a derivation whose component sums are cross-checked in CI, within the 126 that carry a derivation block in all (the spec-v62 A5 named-formula tiles add a substituted-formula line instead of a component sum)). The A3 labeled copy then extended to the six multi-output Group V5 lab / clinical-math tiles (wave 2), the three Group F medication / infusion tiles (wave 3), the three Group I field-medicine tiles (wave 4), and the five Group V7 oxygenation / renal-acid / lipid tiles (wave 5) above — completing the multi-numeric rollout; the remaining hand-built lists (single value plus an interpretation line, unit-conversion / dose-time utilities, and workflow checklists) keep the universal "Copy all".

ICU-infusion, med-surg & OB/neonatal cheat sheet (spec-v62 Part B)

Nine bedside computations across two waves, plus the conversion of the catalog's last two static reference tables into calculators (Part C). Wave 1 shipped seven unambiguous tiles; wave 2 added the two pinned-constant tiles — norepi-equiv (Kotani 2023 norepinephrine-equivalent factors) and neo-phototherapy (AAP-2022 phototherapy + exchange-transfusion curves). This takes the catalog

to 328 (a net +9 from the prior wave). Every tile is a pure lib/clinical-v8.js function, validated through lib/num.js (so a zero/non-finite/out-of-range input throws a caught TypeError/RangeError, never a NaN), fuzz-covered by the spec-v59 object-aware harness, and ships its primary citation inline with a DOI. Every dosing/reversal tile renders the explicit "planning estimate, not an order — verify against local protocol and an independent double-check" notice.

Part C — the last two static tables, now calculators. peds-dose went from a fixed per-kg table to a weight-driven quick-dose panel (each drug computed to actual mg at the entered weight with the per-dose cap applied and flagged), and anticoag-reversal went from an agent table to a weight/INR-driven reversal-dose calculator (4F-PCC Kcentra INR-band dosing with the 100 kg dosing-weight cap, idarucizumab 5 g, andexanet ANNEXA-4, protamine 1 mg/100 units max 50 mg). Both keep their ids and permalinks, both now pass the spec-v29 §3 one-line test — and removing the two lookup-table blocks also retired the last two horizontally-scrolling tables in the catalog, so the 320px no-horizontal-scroll sweep now has no table-scroll region left to guard on a clinical tile.

Part A depth pass — landed in waves on the existing tiles (no count change). A1 (serial/trend mode, lib/trend.js) is wired onto the early-warning family (news2, mews, pews), the hemoglobin-drop tiles (gbs, oakland), and sodium-correction. A2 (the source-anchored "next step" action field, META[id].actions) is seeded on kdigo-aki, ciwa, and cows. A5 (the substituted-formula derivation — the published equation with the user's own numbers plugged in and the arithmetic carried through) is complete across all nine named formula tiles: cockcroft-gault, corrected-sodium, aa-gradient, osmolal-gap, winters, fena-feurea, egfr, drip-rate, and burn-fluid. Each substituted line is guarded at both the author layer (returns null on any missing / non-finite / non-positive input) and the render layer (refuses any string carrying a NaN/Infinity/undefined token), so a bad input can never reach the panel. A4 (the SI⇄conventional lab toggle, the v61 per-field <select> mechanism) is rolling out wave-by-wave: wave 1 wired glucose, BUN, calcium, and albumin toggles onto the Group E correction tiles (corrected-calcium, corrected-sodium, corrected-ca-na, osmolal-gap); wave 2 extended the albumin toggle to the anion-gap and ascites tiles (anion-gap, anion-gap-dd, saag), preserving the optional-albumin empty-check; wave 3 added a magnesium (mg/dL ⇄ mmol/L) toggle to magnesium-replacement; and wave 4 (the final wave) completed the rollout — bilirubin (mg/dL ⇄ µmol/L) on the hepatic and neonatal tiles (meld-childpugh, maddrey-lille, bhutani-bilirubin, psofa, neo-phototherapy), lactate on pelod2, and ionised/total calcium on the three CRRT citrate fields (crrt-dose). The canonical compute unit is always the default option, so every documented example and deep link stays byte-identical (the example-correctness e2e sweep proves it). Lactate and the CRRT calcium fields are SI-canonical (mmol/L is the compute unit, so mmol/L is the default and the conventional mg/dL alternate converts up) — the inverse layout of the conventional-default analytes. A4 is now complete for every lab-input field with a real consumer; phosphate has none (its only candidate, electrolyte-replacement, carries a polymorphic level field whose unit follows the K/Mg/phosphate selector, so a fixed toggle does not fit).

A3 (the reverse-solve / target mode) has shipped on both safety-critical sodium correction tiles: sodium-correction and free-water-deficit now compute a ceiling-capped max-safe rate — when the requested schedule would move Na faster than the published ceiling (8 mEq/L/24 h chronic / 10 acute for raising, 10 for lowering), each surfaces the rate that hits exactly the ceiling and flags it, so the reverse-solve never silently displays an over-ceiling infusion rate. A3 is closed for the qualifying tiles: the spec admits a reverse-solve only where the inverse is single-valued, and the remaining named tiles do not qualify (insulin-drip is an example-only sliding-scale verifier; heparin-nomogram is the Raschke step table, already aPTT-target-seeking; vasopressor/conc-rate already carry the dose⇄rate inverse).

With wave 4 the A4 lab-toggle rollout is complete, and with it spec-v62 Part A is fully shipped (A1 trend, A2 action, A3 reverse-solve, A4 lab toggles, A5 substituted derivation). See docs/spec-v62.md.

Operations depth: deadlines, denial routing & document linting (spec-v63)

The ops surface (billing, coding, regulatory, patient-admin) used to state a rule; it now computes the clock the rule sets, routes a denial, and validates the document against the rule's required elements. v63 has two parts: Part B added five ops calculators (above), and Part A is a zero-tile depth pass that deepens the existing ops tiles the same way spec-v62 deepened the bedside tiles.

The shared primitive — lib/deadline.js (OA1). Before v63 the catalog could compute exactly one regulatory deadline (breach-clock) and had no business-day or federal-holiday math at all. deadline() is pure UTC-midnight arithmetic — no local-timezone drift — over calendar or federal business days:

deadline({ anchor, days, basis, now, rollForward })
   anchor (ISO string | Date)                              now (pin "today")
        │  parseIsoStrict — rejects 2026-13-40                  │
        ▼  (calendar)            (business)                     ▼
   addCalendarDaysUtc      addBusinessDaysUtc ── skip Sat/Sun + 5 U.S.C. 6103
        │                        │   holidays (fixed + floating, federal
        │   rollForward?         │   weekend observance, Dec-31 NYE edge)
        ▼                        ▼
   { deadline, daysElapsed, daysRemaining, pastDue, basis, anchor }

breach-clock was re-pointed onto the engine's date primitives, byte-identical (regression-pinned). Every Part B deadline tile and the OA2 routing run through this one audited path.

Design decision: OA2 and OA3 are decisions and validations, not directories. A denial category is an input the user already has (off their EOB); a completeness finding checks the user's own document. Neither ships anything browsable or searchable, so neither reopens the spec-v29 §3 code/payer-index retirement. And every ops output still carries the regulatory.js posture — it surfaces the regulatory date or level and cites the rule; it never decides whether a breach/overpayment occurred, whether an appeal will succeed, or whether a service is covered.

Calcium replacement: the salt the K/Mg/Phos ladder omits (spec-v64)

The electrolyte-replacement ladder doses potassium, magnesium, and phosphate; calcium is the one electrolyte where the form of the salt is itself the error. Calcium gluconate 10% and calcium chloride 10% are not interchangeable gram-for-gram — and calcium is given in exactly the moments (hyperkalemia, symptomatic hypocalcemia, citrate toxicity, CCB overdose) where the wrong salt or an unnamed "1 g calcium" order is most costly. calcium-replacement (lib/clinical-v7.js calciumReplacement()) computes the confusion away:

So 1 g calcium chloride ≈ 2.94 g calcium gluconate for the same elemental calcium (≈3×). Given a salt + dose, the tile returns the elemental calcium (mg + mEq), the volume, and the equivalent dose of the other salt, plus the standard adult dose for the indication and the precaution notes (slow IV push on a monitor; chloride is sclerosing — central line preferred; never in the same line as bicarbonate or phosphate; caution in digoxin toxicity). Dosing is anchored to AHA ACLS 2020; elemental content to USP / product labeling. It states the dose; it does not write the order. See docs/spec-v64.md.

Bedside physiology a nurse still does on paper (spec-v65)

A render-tree and near-neighbor audit against the live catalog found three deterministic, source-anchored calculations an ICU/ED/floor nurse performs on a routine shift that no existing tile computed. Each passes the spec-v29 §3 one-line test (input → computed output), is a pure lib/clinical-v8.js function fuzz-covered by the spec-v59 harness, and renders the explicit "planning estimate — verify against the device / monitor and local protocol" notice. This takes the catalog to 337.

The three are deliberately distinct from their near-neighbors and cross-linked to them: o2-cylinder-duration is the gas analog of the IV-bag infusion-time-remaining; minute-ventilation is the gas-exchange calc the mechanics tiles (driving-pressure, pbw-ardsnet, rsbi) do not cover; and cerebral-perfusion-pressure extends map (which computes MAP from blood pressure but never subtracts ICP). Cylinder factors (D 0.16 / E 0.28 / M 1.56 / G 2.41 / H-K 3.14 L/psi) are physical constants of the cylinder geometry; the ventilation math is anchored to Marino's ICU Book; CPP and its target band to the Brain Trauma Foundation 2017 guideline (Carney N, Neurosurgery 2017;80(1):6-15). A gauge at/below the safe residual flags "swap now" rather than rendering a negative duration, and a negative CPP (ICP > MAP) is surfaced with an explicit critical-low flag, never hidden. See docs/spec-v65.md.

Cardiology & ECG cheat sheet (spec-v90, Wave 2 of the spec-v85 program)

Six deterministic cardiology / ECG computations that fill confirmed gaps in the catalog's cardiology surface (it had qtc-suite, sgarbossa, map, and the spec-v87 hemodynamics, but none of these six). Each passes the spec-v29 §3 one-line test, is a pure lib/cardio-v90.js function fuzz-covered by the spec-v59 harness, and quotes the cited source's own band — none reads a waveform or auto-disposes. This takes the catalog to 385.

The two ill-defined inputs are domain-guarded so no non-finite value reaches the DOM (spec-v59): ecg-axis surfaces the all-isoelectric (0,0) complex as an "indeterminate axis" rather than a spurious 0° or NaN, and aortic-valve-area guards division by AV_VTI = 0. Five are Class A fixed instruments (hexaxial geometry; the 1949/1985 voltage thresholds; the Morrow 2000 weights; the Mark 1987 coefficients; the constant 451) with no staleness row; aortic-valve-area is Class B — its ASE/EACVI 2017 + 2020 ACC/AHA severity cutoffs carry a citation-staleness row and are the first subject of the new scripts/check-citation-cadence.mjs warn-only monthly job. See docs/spec-v90.md.

Pulmonary function & chronic respiratory disease cheat sheet (spec-v91, Wave 2 of the spec-v85 program)

Five deterministic pulmonary computations that fill the chronic gap beside the catalog's acute respiratory surface (aa-pf-suite, rox, curb-65, smart-cop). These are the PFT-lab and the COPD/ILD-clinic standards: the GOLD spirometric grade, the BODE COPD prognosis, the GAP index for IPF, the GLI-2012 predicted-spirometry reference, and the mMRC dyspnea scale that feeds the first two. Each passes the spec-v29 §3 one-line test, is a pure lib/pulm-v91.js function fuzz-covered by the spec-v59 harness, and quotes the cited source's own grade / band / mortality. This takes the catalog to 390.

Grade inputs are clamped to their published range (mmrc-dyspnea and the mMRC that bode-index consumes accept only 0–4; gap-ipf treats "cannot perform" DLCO as a first-class 3-point state, never a blank). gold-spirometry and predicted-spirometry guard every division/ln domain so a zero/blank input surfaces "(complete the fields)" rather than a NaN/Infinity. The GLI-2012 coefficient + spline sets are compiled module constants (lib/gli-2012-data.js, spec-v85 §5 — not a data/ dataset), transcribed from the published GLI lookup table; the 40-yr/175-cm Caucasian-male predicted FEV1 (4.08 L) / FVC (5.05 L) / FEV1/FVC (0.81) and their LLNs reproduce the published reference values. Three tiles are Class A fixed instruments (Celli 2004, Ley 2012, Bestall 1999); gold-spirometry (GOLD 2024, annual) and predicted-spirometry (GLI-2012, on-publication) are Class B and carry citation-staleness rows read by scripts/check-citation-cadence.mjs. See docs/spec-v91.md.

Nephrology: CKD staging, proteinuria ratios, dialysis adequacy, contrast-nephropathy risk & cystatin-C eGFR cheat sheet (spec-v92, Wave 2 of the spec-v85 program)

Five deterministic nephrology computations that close the chronic / procedural renal gap beside the catalog's existing filtration / injury / dosing surface (egfr-suite, fena-feurea, kdigo-aki, cockcroft-gault). These are the nephrology-clinic and dialysis-unit standards: the KDIGO G×A risk heat-map, the spot albumin/protein ratios, hemodialysis adequacy, the Mehran contrast-risk score, and the race-free cystatin-C eGFR. Each passes the spec-v29 §3 one-line test, is a pure lib/nephro-v92.js function fuzz-covered by the spec-v59 harness, and quotes the cited source's own cell / ratio / target / band / estimate. This takes the catalog to 395.

The two-axis ckd-staging is band-mapping over ordinal axes (it accepts the A-category directly when no numeric UACR is given and never emits an unlabeled cell), and uacr-upcr shares its A-stage cutoffs so the two agree. Every division (uacr-upcr urine creatinine, ktv-urr pre-BUN), logarithm (ktv-urr domain R − 0.008·t > 0), and power term (ckd-epi-cystatin cystatin/creatinine bases) is domain-guarded so a zero/blank input surfaces a labeled fallback rather than a NaN/Infinity. Four tiles are Class A fixed instruments (the ratio math, the Daugirdas Kt/V, the 2004 Mehran weights, the 2021 CKD-EPI coefficients); ckd-staging (KDIGO 2024, on-publication) is Class B and carries a citation-staleness row read by scripts/check-citation-cadence.mjs. See docs/spec-v92.md.

Hepatology & GI disease activity: NAFLD fibrosis, Glasgow-Imrie pancreatitis, Truelove-Witts, Harvey-Bradshaw, Mayo UC & Milan criteria cheat sheet (spec-v93, Wave 2 of the spec-v85 program)

Six deterministic hepatology & GI disease-activity instruments that close the catalog's liver/gut gap beside the existing chronic-liver and pancreatitis spine (meld-childpugh, fib4, apri, ranson-bisap, maddrey-lille). These are the disease-activity and fibrosis instruments a hepatology and GI clinic score constantly: the NAFLD-specific non-invasive fibrosis estimate, the parallel UK/European pancreatitis severity score, the two ulcerative-colitis activity indices, the Crohn's index, and the HCC transplant-eligibility criterion. Each passes the spec-v29 §3 one-line test, is a pure lib/hepgi-v93.js function fuzz-covered by the spec-v59 harness, and quotes the cited source's own band / class / index / criterion. This takes the catalog to 401.

The one guarded domain is nafld-fibrosis's AST/ALT division — a blank/zero ALT surfaces a labeled fallback rather than a NaN/Infinity term; the other five are point-table or decision logic. glasgow-imrie never lets a partial 48-hour panel masquerade as a complete low score, truelove-witts reports a near-miss as the band it actually falls in, mayo-uc keys the partial-vs-full fallback on the endoscopy subscore, and milan-criteria guards a zero count / missing size. All six are Class A fixed published derivations (Angulo 2007, Blamey/Imrie 1984, Truelove & Witts 1955, Harvey-Bradshaw 1980, Schroeder 1987, Mazzaferro 1996), so none carries a citation-staleness row — their citations name journals and authors, not a recurring guideline issuer. See docs/spec-v93.md.

Hematology & oncology prognosis: HScore, IPSS-R, FLIPI/IPI, MASCC & Sokal/ELTS cheat sheet (spec-v94, Wave 2 of the spec-v85 program)

Five deterministic heme/onc prognostic scores that close the catalog's malignancy-prognosis gap beside the existing heme bedside cluster (anc, khorana, four-ts, isth-dic, tls-cairo-bishop). These are the scores an oncologist or hematologist computes to stratify a new diagnosis and set the survival expectation: the weighted diagnostic score for reactive HLH, the MDS prognosis index, the two lymphoma five-factor indices, the febrile-neutropenia disposition index, and the two at-diagnosis CML risk formulas. Each passes the spec-v29 §3 one-line test, is a pure lib/hemonc-v94.js function fuzz-covered by the spec-v59 harness, and quotes the cited source's own band / category / index. This takes the catalog to 406.

The load-bearing guarded domains are in sokal-cml: the ELTS (platelets/1000)^−0.5 term divides by the platelet count (a zero/negative platelet surfaces a labeled fallback), and the Sokal exp() overflows to Infinity for an extreme age/platelet input — surfaced as a finite null, never an Infinity term. The other four are point-table logic; mascc reports the index only, not the admission decision, and ipss-r-mds ships the clinical/cytogenetic IPSS-R, not the molecular IPSS-M. All five are Class A fixed published derivations (Fardet 2014, Greenberg 2012, Solal-Céligny 2004 / IPI 1993, Klastersky 2000, Sokal 1984 / Pfirrmann 2016), so none carries a citation-staleness row. See docs/spec-v94.md.

Neurology outcome & grading: modified Rankin, GOS-E, Hoehn-Yahr, Spetzler-Martin, House-Brackmann & MIDAS cheat sheet (spec-v95, Wave 2 of the spec-v85 program)

The catalog's neurology surface was acute-onset, not longitudinal: a clinician could compute the NIHSS at presentation, the ICH 30-day mortality risk, the SAH grade, the coma score, and the post-TIA stroke risk — but nothing for the next visit. These six fill that gap: the stroke-trial functional-outcome endpoint, the TBI outcome at six months, the Parkinson stage at clinic, the AVM surgical-risk grade, the facial-nerve recovery grade, and the migraine-disability band. They are ordinal selectors and bounded-integer sums (no division, root, or log except none at all), pure lib/neuro-v95.js functions fuzz-covered by the spec-v59 harness, each quoting the cited source's own descriptor and band.

mrs, gose, hoehn-yahr, and house-brackmann are arithmetic-free ordinal selectors — an out-of-range or blank selection surfaces a labeled valid:false fallback, never a wrong band; the GOS-E↔GOS map is validated both directions. spetzler-martin clamps the core grade to 1–5 and the supplemented total to 2–10 by construction and surfaces the derivation; midas coerces blanks to 0, clamps each day-count to the 92-day window, and excludes the ancillary items from the sum. All six are Class A fixed ordinal definitions (van Swieten 1988, Wilson 1998, Hoehn-Yahr 1967, Spetzler-Martin 1986 / Lawton-Young 2010, House-Brackmann 1985, Stewart 2001), so none carries a citation-staleness row. See docs/spec-v95.md.

Psychiatry: the clinician-rated severity scales one rung above the screeners (spec-v96, +6 → 418)

The catalog already carried the brief, validated self-report screeners a nurse or primary-care clinician hands a patient: phq9 and gad7, cssrs for suicide risk, gds15, epds, and auditc. What it had no tile for is the layer above the screen — the clinician-rated rating scales that measure severity and track change on treatment. A phq9 is what the patient says; the HAM-D is what the clinician rates. spec-v96 ships six, all Group G, pure lib/psych-v96.js functions fuzz-covered by the spec-v59 harness.

The summed scales (hamd/hama/madrs/ybocs/pcl5) refuse a band from a partially-completed instrument (spec-v57): a blank item renders "(complete all N items)" and no band — an unanswered item is not a zero — and an out-of-range item yields a labeled valid:false rather than a silently-wrong sum. mdq is a fixed three-gate rule, never positive on the symptom count alone, and the pcl5 cutoff is quoted as the published range, not a single hard threshold the catalog invents. All six are Class A (fixed published item weights and author-defined bands: Hamilton 1959/1960, Montgomery-Åsberg 1979, Hirschfeld 2000, Goodman 1989, Blevins 2015), so none carries a citation-staleness row. The copyrighted/licensed instruments (MoCA, SLUMS, BDI-II) are excluded for licensing. See docs/spec-v96.md.

Perioperative risk: the probability and the score, one rung above the screen (spec-v97, +5 → 423)

The catalog's pre-op surface was strong on the screening indices a clinician runs in clinic — rcri (a Lee class), ariscat (pulmonary-complication risk), lemon (a difficult-airway screen), apfel (PONV), plus the spec-v89 asa-ps and surgical-apgar. What it lacked is the layer above: the published regression equations that return an actual predicted probability rather than a risk class, and the validated weighted indices an anesthesiologist reaches for on a high-stakes case. spec-v97 ships five, all Group G, pure lib/periop-v97.js functions fuzz-covered by the spec-v59 harness.

Two design points keep these honest. The Gupta logistic link is overflow-guarded: the linear predictor x is clamped to [−40, 40] before e^−x, so even a fuzzed 1e9 age returns a finite probability in [0, 100], never NaN/Infinity — and every categorical input is validated against its fixed enum (an out-of-enum value surfaces valid:false, never a silent NaN). The point tables are transcribed, not approximated: the Gupta coefficients were cross-checked against two independent reproductions of the source models, and the entire POSPOM age / 15-comorbidity / 24-procedure point system and its points→mortality lookup were transcribed verbatim from the paper's Supplemental Digital Content 3 and spot-verified against the source. All five are Class A (fixed regression coefficients / published point tables: Gupta 2011 ×2, Arozullah 2001, el-Ganzouri 1996, Le Manach 2016), so none carries a citation-staleness row. The proprietary ACS-NSQIP universal Surgical Risk Calculator is excluded — it is a hosted model, not a fixed published equation. See docs/spec-v97.md.

Pediatrics: the four standard rules the deep neonatal surface still lacked (spec-v98, +4 → 427)

The catalog's pediatric surface was already broad — Group N carries the neonatal and procedural tiles (ballard, finnegan, bhutani-bilirubin, downes, neo-phototherapy, pecarn-head, pecarn-cspine, pecarn-iai), and Group G the pediatric clinical scores (pews, peds-gcs, alvarado-pas, nigrovic, the febrile-infant rules, westley, pram-asthma, pelod2, psofa). A full-catalog sweep (the first draft proposed five tiles that turned out already shipped) left exactly four genuinely-absent standard instruments. spec-v98 ships them, all Group G, pure lib/peds-v98.js functions fuzz-covered by the spec-v59 harness.

pim3 uses the published Straney 2013 coefficients — not the PIM3-anz13 registry recalibration that also circulates — cross-verified against two independent reproductions; its logistic is clamped before e^−x so a fuzzed 1e9 input still returns a finite probability in [0, 100]. kawasaki-criteria is Class B (the AHA statement is revisable → a citation-staleness row); the other three are Class A. No growth-percentile chart tile (a dataset, out per spec-v29 §3); no auto-CT / auto-aspiration order. See docs/spec-v98.md.

ID, critical care & burns: closing the spec-v85 program (spec-v99, +5 → 432)

The catalog had the acute-infection and critical-care triage tools (curb-65, sirs, qsofa-sofa, smart-cop, apache2) and the burn-resuscitation calculator (burn-fluid, which takes %TBSA as an input), but five standard ID/critical-care/burns instruments were absent. spec-v99 ships them — the tenth and final feature spec of Wave 2 — all Group G, pure lib/idcrit-v99.js functions fuzz-covered by the spec-v59 harness, closing the spec-v85 Advanced Clinical Calculators program at 432 tiles (+66 across the ten feature specs v86 through v99).

Two correctness anchors. saps-ii is transcribed and calibration-checked: the 17-variable point bands were cross-verified against MDCalc and ClinCalc (a corrupted "+1" urine-output band in one reproduction was rejected in favor of the correct "+4"), and the worked 64-point case → 75.3% matches the published ClinCalc calibration; the mortality logistic and ln(SAPS+1) are domain-guarded. lund-browder sums to exactly 100% at every age band (cross-verified against the Joint Trauma System adult/pediatric charts); region fractions clamp to [0, 1] and a > 100% total is flagged, not silently capped. duke-endocarditis (2023 Duke-ISCVID) and refeeding-risk (NICE CG32) are Class B with citation-staleness rows; the other three are Class A. See docs/spec-v99.md.

Advanced Bedside Quantitation program (spec-v185, v186, v187, +19 → 793)

Three post-audit feature specs add 19 genuinely-missing calculators. An earlier draft batch had proposed roughly forty calculators via a faulty keyword scan; most were already shipped, so those drafts were withdrawn and the work re-done honestly: every id below was verified absent by a direct scan of app.js before any code was written, and every constant re-fetched and cross-verified against ≥ 2 independent sources at implementation (spec-v97).

v185 closes the invasive/echo hemodynamics set (Fick cardiac output with a LaFarge-estimated VO₂, the Gorlin valve-area equation, the Qp/Qs shunt ratio, and Doppler LVOT-VTI stroke volume), two anticoagulation-quality tools (VTE-BLEED and the Rosendaal time-in-therapeutic-range over a dated INR series), the Matsuda OGTT insulin-sensitivity index, and Janmahasatian lean body weight. v186 adds radiotherapy BED/EQD2, PISA regurgitant-orifice quantification, LV meridional wall stress, the hemoglobin-corrected DLCO (Cotes), estimated VO₂max/METs (Bruce/Cooper), and the Wilson-score proportion CI. v187 opens the Subspecialty Oncology & Hematology Staging program with BCLC, IMDC/MSKCC metastatic-RCC risk, RECIST 1.1, and the modified Glasgow Prognostic Score. Each compute is finite-/positive-guarded and covered by the spec-v59 fuzz harness; dlco-correction carries a documentation-only citation-staleness row (ATS trips the issuer-acronym pattern; the Cotes formula is unchanged).

Hepatology/GI, Dermatology/Urology & Screening/Risk program (spec-v190, v191, v192, +12 → 814)

Three feature specs add 12 deterministic calculators across three under-represented surfaces, each id verified absent by a direct app.js scan before any code (spec-v85 §6.2) and each coefficient, boundary, point weight, and criterion re-fetched and cross-verified against ≥ 2 independent sources at implementation (spec-v97).

v190 fills the hepatology/GI gap: the PALBI grade (the platelet-augmented ALBI, using the published log-quadratic linear predictor and the −2.53 / −2.09 cut-points), MELD-Na (the sodium-augmented MELD in its OPTN/UNOS operational form — sodium applied only when MELD > 11, bounded 6–40, distinct from the Kim-2008 NEJM re-fit), the Clichy acute-liver-failure criteria (encephalopathy plus an age-30-branched factor-V threshold), and the Rome IV IBS criteria with the IBS-C/D/M/U subtype. v191 spans dermatology and urology: SCORTEN (the toxic-epidermal-necrolysis mortality bands), the AJCC 8th-edition melanoma T category (the 0.8 mm split and ulceration a/b suffix, T element only), PI-RADS v2.1 (the zone-specific score-3 upgrade rules — peripheral DWI-3 + DCE, transition T2W-3 + DWI-5), and the Guy's stone score (PCNL complexity Grade I–IV). v192 adds bedside risk: FINDRISC (type-2-diabetes screening), the Grobman race-free 2021 VBAC calculator (the published logistic model, which uses weight + height rather than BMI, evaluated in odds space per spec-v140), the Marburg Heart Score (rule out CAD in primary-care chest pain), and the ADHERE in-hospital heart-failure mortality CART tree. The fifth proposed v192 score, GWTG-HF, was deferred: its complete row-by-row sub-range point table (Peterson 2010 Table 3) is paywalled and not reproduced verbatim in ≥ 2 independent open sources, so it is parked with precise-dapt under the spec-v97 fidelity bar rather than shipped from a continuous-variable approximation. Every compute is finite-guarded and covered by the spec-v59 fuzz harness (the grobman-vbac logistic in odds space with a [0, 1] clamp); no tile trips the issuer-acronym pattern (AJCC / ACR are not in it), so none forces a staleness row.

Subspecialty Oncology & Hematology Staging closeout (spec-v188, v189, +9 → 802)

Two feature specs close the program v187 opened with 9 more staging / prognostic instruments, each id verified absent by a direct app.js scan before any code (spec-v85 §6.2) and each weight re-fetched and cross-verified against ≥ 2 independent sources (spec-v97).

v188 ships the two CLL clinical stages (Binet, Rai), Ann Arbor / Lugano lymphoma staging, FLIPI-2 (the β₂-microglobulin revision, kept distinct from the live FLIPI-1), and the Hasford (Euro) CML score beside the live Sokal. v189 extends into adjacent prognosis: mSMART myeloma cytogenetic risk (with double/triple-hit naming), the IMPEDE-VTE myeloma thromboprophylaxis score, SAMe-TT2R2 for VKA-control prediction, and the Elixhauser comorbidity index under the original van Walraven (2009) signed weighting (range −7 to +12) as a complement to the live Charlson. The fifth proposed v189 score, BVAS v3, was deferred: a faithful score needs item-level new/worse-vs-persistent weighting of ~56 items across nine organ systems, and an organ-system approximation would misreport the total — parked with precise-dapt under the spec-v97 fidelity bar rather than shipped from an approximation. Every compute is finite-guarded, takes only bounded comparisons or integer sums, and is covered by the spec-v59 fuzz harness; no tile trips the issuer-acronym pattern, so none forces a staleness row.

Long-Term Care & Geriatric Assessment program: older-adult mortality & LTC prognosis (spec-v172, v180, +2 → 816; 5 deferred)

spec-v180 opens cluster §3.8 of the LTC-GA program — the validated life-expectancy and health-instability instruments the nursing home, the geriatric clinic, and the hospice team reach for when deciding whether to screen, to refer, or to revisit the goals of care. It ships 2 of its 7 proposed tiles (both Clinical Scoring & Risk, Group G); the remaining five are deferred on the spec-v97 ≥ 2-source verbatim bar:

• lee-mortality-index — the Lee 4-Year Mortality Index for older adults (Lee, JAMA 2006). A weighted point sum (0–26): an age band (60–64 = 1 up to ≥ 85 = 7), male sex (2), diabetes (1), cancer (2), chronic lung disease (2), heart failure (2), current smoking (2), BMI < 25 (1), and difficulty bathing (2), walking several blocks (2), managing money (2), pushing/pulling heavy objects (1). The total maps by table lookup to the validation-cohort 4-year all-cause mortality bands (0–5 ≈ 4%, 6–9 ≈ 15%, 10–13 ≈ 42%, ≥ 14 ≈ 64%). Because it is a published point-total → observed-mortality table, there is no exponentiation and no 1 − sigmoid(−bx) complement — the spec-v140 saturation hazard cannot arise.

• chess-scale — the interRAI CHESS scale (Changes in Health, End-stage disease, Signs and Symptoms; Hirdes, J Am Geriatr Soc 2003), operationalized per the interRAI LTCF Outcome Scales (CIHI). Count the signs/symptoms present (vomiting, edema, dyspnea, weight loss, dehydration/low fluid, reduced intake), capped at 2, then add one point each for decline in decision-making, decline in ADL status, and an end-stage (≤ 6-month) prognosis — a 0–5 health-instability score.

Both are prognostic estimates framed as decision support for life-expectancy-informed care planning, never a prediction of an individual's death and never an end-of-life order in Sophie's voice (spec-v11 §5.3). Both are Class A journal formulas naming no ISSUER_PATTERN acronym, so neither takes a citation-staleness row. Every weight, band, item, and combination rule was re-fetched and cross-verified against ≥ 2 independent sources (spec-v97): Lee across the JAMA Table 3/4, the abstract / MDCalc reproduction, and the SoFOG "Score de Lee" PDF; CHESS across the interRAI official CHESS PDF, the CIHI interRAI LTCF Outcome Scales Reference Guide (with a worked example scoring 4/5), and the CIHI interRAI Contact Assessment job aid. Deferred (each re-opens when it clears the bar): schonberg-index (5-year weights double-sourced, but the point→mortality-band percentages are single-sourced and the 9-year weights are published only as a non-extractable image), walter-index, suemoto-index, mitchell-mri, and adept. New lib/ltcga-v180.js + views/group-v180.js (RV180). See docs/spec-v180.md.

Deep Subspecialty Quantitation program (spec-v199–v203, +17 so far → 891)

The deepest specialist stratum of the scope-mdcalc-parity tail: the myeloid-malignancy prognostic scores a hematologist computes at diagnosis, the intensivist's whole-database severity models, the hepatologist's decompensation math, the cardiologist's multivariable survival engines, and the perioperative / frailty instruments that gate a surgical or geriatric decision. The program opens with spec-v199 and its myeloid-neoplasm & transplant slice.

Design decisions. (1) Absence is a functional test, not an id test. The proposed fifth tile (ELTS) was dropped at implementation — the spec-v85 §6.2 collision re-check found the EUTOS Long-Term Survival score already computed by the live sokal-cml tile (lib/hemonc-v94.js), with identical coefficients and bands, so a standalone tile would duplicate it. The program opens +4, not +5. (2) Verify against the source, correct the draft. The spec-v97 ≥ 2-source re-verification caught two draft slips: the HCT-CI rheumatologic and peptic-ulcer weights are +2 each (Sorror 2005 / MDCalc), not +1; and the GIPSS range is 0–6, not 0–8. (3) Grade the mutually-exclusive severities with a select. MIPSS70's HMR count and the HCT-CI hepatic / pulmonary rows are single <select>s so an impossible double-count (e.g. "1 HMR" and "≥ 2 HMR") is unrepresentable. All four are decision support, never a transplant / conditioning / chemotherapy order (spec-v11 §5.3).

v200 design decisions. (1) Absence is a functional test, again. The proposed fifth tile (a vasoactive-inotropic-score tile) was dropped at implementation — the spec-v85 §6.2 collision re-check found VIS already computed by the live vis tile (lib/clinical-v4.js, spec-v13) with the identical Gaies 2010 multipliers, so v200 opens +4, not +5 (the ELTS precedent). (2) Verify against the source, correct the draft. The spec-v97 re-verification caught three APPS drafting slips against Villar 2016: the PaO₂/FiO₂ middle band is 105–158 (not 84–158), the plateau middle band is > 27–30 (not 28–29), and the mortality tiers are 5–7 / 8–9 (not 5–6 / 7–9). (3) Grids doubly-verified, mortality models sourced honestly. The OASIS and LODS point grids were transcribed band-for-band from two independent open reproductions each; their logistic mortality coefficients rest on one open source (author reference code / a Le-Gall-1996 reproduction), literature-corroborated and sanity-checked, and are presented as model estimates. All four are decision support, never a titration / ventilator / fluid / disposition order.

Advanced Specialist Quantitation program (spec-v193–v198, +28 → 844)

The remaining scope-mdcalc-parity tail is no longer the common bedside scores — those are carried — but the specialist-grade instruments a subspecialist reaches for: the interventional cardiologist's post-PCI risk models, the intensivist's invasive-hemodynamic and gas-exchange math, the hepatologist's transplant-free-survival models, the endocrinologist's thyroid-homeostasis and β-cell indices, and the subspecialty prognostic scores in heme-onc, ID, neurology, and gyn-onc. Six feature specs ship 28 deterministic, cited, actionable instruments across six new lib modules and six renderers. Every point weight, coefficient, and band threshold was re-fetched and cross-verified against ≥ 2 independent open sources at implementation (spec-v97); each compute routes through lib/num.js, is finite-guarded at its zero-denominator / log-domain edges, and is covered by the spec-v59 fuzz harness with zero non-finite leaks. All are decision support, never an order (spec-v11 §5.3).

Design decisions. (1) Verbatim over recall. The coefficient-heavy models (GLOBE, UK-PBC, SPINA, the CRUSADE point grid, the FIGO 0/1/2/4 table) were transcribed from primary sources and reproduced against a second — the SPINA-GD factor 601 = 1 + K₃₀·[TBG] was the one place a plausible simplified form is wrong by ~600×, caught by reproducing the package's own worked example. (2) Population context travels with the number. PAPi renders both the acute-MI (< 1.0) and advanced-HF (< 1.85) thresholds; the S/F and OSI tiles render the SpO₂ ≤ 97% ceiling caveat in-tile, so a value is never read outside its validity range. (3) One issuer row. Only figo-gtn names "WHO" in its citation, so it takes the sole new citation-staleness row; the other 27 name journals and authors and stay documentation-free. Deferred on the spec-v97 reproducibility bar: GRACE 2.0 (restricted-cubic-spline coefficients not openly published), VOCAL-Penn, HOMA2 (closed iterative model), and GWTG-HF / EMSE / CRASH-TBI (paywalled or single-source point tables). New lib/acs-v193.js–subspecialty-v198.js + views/group-v193.js–group-v198.js (RV193–RV198).

Long-Term Care & Geriatric Assessment program: infection surveillance & antimicrobial stewardship (spec-v172, v181, +2 → 774)

spec-v181 closes cluster §3.9 of the LTC-GA program with the tight, high-value pair a nursing home's CMS-mandated infection-prevention-and-control program (IPCP) and antibiotic-stewardship program (ASP) run on:

• mcgeer-criteria — the Revised McGeer surveillance definitions (Stone 2012). Pick the suspected site, check the constitutional + site-specific findings → MEETS / DOES NOT MEET the surveillance definition, naming the satisfied criteria and the blocking gap. Ships the cross-verified syndromes: UTI (with / without catheter), respiratory (common cold/pharyngitis, influenza-like illness, pneumonia, lower-RTI), skin & soft tissue (cellulitis/ wound, conjunctivitis), and gastroenteritis. A surveillance definition for tracking and reporting — not a diagnosis and not a treatment trigger.

• loeb-minimum-criteria — the Loeb minimum criteria for initiating antibiotics (2001). Pick the site → minimum criteria MET / NOT MET across UTI (with / without catheter), lower respiratory (all 5 Loeb paths), skin & soft tissue, and fever-of-unknown-source. Stewardship decision support — it neither orders nor withholds antibiotics and names no agent, dose, route, or duration; the prescriber and local protocol decide.

Both are categorical, site-branched criteria-logic determinations (no numeric score, no numeric leak — fuzzed for the empty/partial-selection and false-positive paths). Every criterion, body-site definition, temperature threshold, and boolean rule was re-fetched and cross-verified verbatim against ≥ 2 independent sources (spec-v97): the Stone 2012 primary paper plus the Missouri DHSS and Minnesota DOH field tools (McGeer), and the MN DOH card plus MO DHSS chart (Loeb). Both are Class A — the journal citations name no ISSUER_PATTERN acronym (SHEA is not in the pattern), so no citation-staleness row. The Stone 2012 systemic primary-bloodstream / unexplained-febrile definitions and the rash-plus-provider-diagnosis dermatologic sub-syndromes are deferred on sourcing / computability grounds. New lib/ltcga-v181.js + views/group-v181.js (RV181). See docs/spec-v181.md.

Long-Term Care & Geriatric Assessment program: cognition & dementia staging (spec-v172, v173, +3 → 740)

spec-v172 opens the fifth-pass program — the nursing-home / skilled-nursing / hospice surface the acute-care passes never indexed (the MDS 3.0, interRAI, and dementia-staging instrument families). spec-v173 is its first feature spec; it ships the three cognition / dementia-staging tiles whose exact item-level scoring could be re-fetched and cross-verified against ≥ 2 independent sources (spec-v97), and defers the other five of the eight proposed until their scoring is verbatim-verifiable — the same safety doctrine that defers a data-sourced tile when the source can't be fetched.

All three shipped tiles are Group G, Class A, trip no ISSUER_PATTERN (CMS/MDS method and journal issuers are not in it), and carry ≥ 3 boundary worked examples with band-flips (BIMS 7→8, AD8 1→2, CDR-SOB 4.0→4.5 and 9.0→9.5). The compute lives in lib/ltcga-v173.js (fuzzed, zero non-finite leaks); the deferrals are recorded in spec-v173 and the parity ledger.

Long-Term Care & Geriatric Assessment program: continence, caregiver strain & advanced wound (spec-v172, v182, +5 → 772; Waterlow deferred)

spec-v182 closes the implemented portion of the LTC-GA program. It completes three high-traffic nursing-home surfaces the acute-care passes never indexed — continence severity, caregiver strain, and advanced wound assessment — shipping 5 of 6:

The two caregiver instruments are the free alternatives to the licensed Zarit Burden Interview (excluded by design); bwat is the full healing-trajectory companion to the live braden / norton-push pressure-injury tiles. waterlow is deferred: the Waterlow card has detailed per-category sub-weights with documented edition drift (1985 vs the 2005 revised card), and the current-card table could not be byte-verified against ≥ 2 open sources — the same sourcing gate applied throughout this program. The compute lives in lib/ltcga-v182.js (fuzzed, zero non-finite leaks).

Program status. The LTC-GA program (v173–v182) is implemented except spec-v180 (older-adult mortality indices — Lee, Schonberg, Walter, Suemoto, Mitchell MRI, ADEPT, CHESS) and spec-v181 (LTC infection-surveillance criteria — revised McGeer, Loeb). Both remain open by deliberate choice: each turns on high-stakes content (mortality-prediction point tables; antibiotic-initiation per-site boolean logic) that must be transcribed verbatim from primary sources to meet the spec-v97 bar, and that verbatim sourcing is the work a future session will do before shipping them.

Long-Term Care & Geriatric Assessment program: polypharmacy burden (spec-v172, v179, +3 → 767; MRCI deferred)

spec-v179 is the program's seventh feature spec (cluster §3.7). The live beers-check flags individual inappropriate medications; v179 adds the cumulative-burden view a deprescribing review needs. The design point worth recording is the spec-v100 §2 classification clarification: these scales are published as per-drug lookup tables, but a lookup table is a reference card, not a calculator (it fails spec-v29 §3). So none of these tiles embeds the drug database — each consumes the per-drug inputs the clinician reads from the published scale and does the arithmetic:

drug-burden-index is a guarded sum of ratios: δ must be finite and positive, so a blank or partial drug row or a zero δ returns a surfaced valid:false rather than Infinity, and the division path is fuzzed (spec-v59). medication-regimen- complexity (MRCI) is deferred: its 65-item Section A/B/C weight tables (George 2004) are paywalled and copyright, and could not be byte-verified against ≥ 2 open sources at implementation — the same sourcing gate that governed the deferrals in the cognition, frailty, and polypharmacy specs. The compute lives in lib/ltcga-v179.js (fuzzed, zero non-finite leaks).

Long-Term Care & Geriatric Assessment program: geriatric nutrition & dysphagia (spec-v172, v178, +6 → 764)

spec-v178 is the program's sixth feature spec (cluster §3.6). The live nutrition screens (must-nutrition, nrs2002, mnutric) are admission / ICU triage tools; the long-term-care home runs on a different set — the lab-based geriatric indices, an appetite screen, a patient dysphagia self-report, and the community-elder checklist. v178 ships all six:

The three lab indices are guarded formulas in Group E: gnri divides body weight by Lorentz ideal body weight, and that denominator is positive-checked so a degenerate height returns a surfaced valid:false rather than Infinity (the division path is fuzzed). The determine weights were not taken from memory — they were pulled verbatim from the ACL Nutrition Screening Initiative checklist (pdftotext over the official PDF) and they sum to exactly 21, the published maximum. eat-10 is cross-linked to the live guss clinician swallow test as its patient self-report complement; snaq carries an inline disambiguation from the similarly named Short Nutritional Assessment Questionnaire. The compute lives in lib/ltcga-v178.js (fuzzed, zero non-finite leaks).

Long-Term Care & Geriatric Assessment program: frailty & sarcopenia case-finders (spec-v172, v177, +4 → 758; 3 deferred)

spec-v177 is the program's fifth feature spec (cluster §3.5). The live frailty surface (frail-scale, mfi-5/11, ves-13) is deficit-count / self-report; v177 adds the sarcopenia case-finders and the multidomain LTC frailty screens. Four ship; three are deferred — and the deferrals are the point worth recording, because they show the sourcing bar holding:

The Rockwood Clinical Frailty Scale is copyright Dalhousie University and requires a license for commercial use; reproducing its nine anchored descriptors fails the free-reproducibility bar (the same call as PACSLAC in v175). The GFI and EFS were deferred for a different reason: their exact per-item 0/1 and 0/1/2 thresholds (the GFI fitness-rating cut, the EFS domain points summing to 17) could not be byte-verified against ≥ 2 independent sources at implementation, so they fall under the spec-v97 sourcing gate rather than ship from an approximate recall — the same discipline that deferred 5 tiles in v173. The four shipped tiles live in lib/ltcga-v177.js (fuzzed, zero non-finite leaks); prisma-7 carries the reverse-scored support item as a unit-tested edge.

Long-Term Care & Geriatric Assessment program: falls-risk, balance & gait (spec-v172, v176, +6 → 754)

spec-v176 is the program's fourth feature spec (cluster §3.4). The catalog carried the two inpatient falls-risk screens (morse-falls, hendrich-ii); it lacked the performance-based battery and the community / LTC screening algorithm a nursing home, geriatric clinic, or outpatient PT uses. v176 ships all six, every norm and cut-point re-fetched and cross-verified against ≥ 2 independent sources (spec-v97):

Two engineering points are worth recording. gait-speed is a guarded ratio: the time denominator is finite/positive-checked, so a zero or blank time returns a surfaced valid:false rather than Infinity — and the division path is explicitly fuzzed (spec-v59). And chair-stand-30s / functional-reach refuse to guess: an age outside the normed strata returns valid:false, never a fabricated band. The three CDC-STEADI-derived tiles trip the ISSUER_PATTERN on “CDC” and so are Class B, each carrying a docs/citation-staleness.md row. gait-speed is the only tile here in Group E (it returns a value, m/s). The compute lives in lib/ltcga-v176.js (fuzzed, zero non-finite leaks). berg-balance, tinetti-poma, and tug are reserved to other specs and are not shipped here.

Long-Term Care & Geriatric Assessment program: observational pain in the cognitively impaired elder (spec-v172, v175, +3 → 748)

spec-v175 is the program's third feature spec (cluster §3.3). The catalog already carried painad and cpot — strong observational pain tiles for the nonverbal patient — but an LTC pain protocol frequently mandates a specific instrument, and a facility whose policy names Abbey or DOLOPLUS-2 cannot substitute PAINAD. v175 ships all three mandated scales, each item list, per-item range, and band re-fetched and cross-verified against ≥ 2 independent sources (spec-v97):

The instrument worth its own note is cnpi: unlike painad it structures the with-movement assessment as a separate condition, so the compute carries two independent 0–6 sums (rest, movement) and a 0–12 combined total and never scores movement from rest — a blank condition renders a complete-the-fields fallback. abbey-pain is the standard scale in Australian and UK aged care; doloplus-2 the standard in French and European geriatric care. All three are Group G, Class A, trip no ISSUER_PATTERN, and carry band-flip boundary worked examples (Abbey 7→8 and 13→14, DOLOPLUS-2 4→5, CNPI rest-vs-movement split). The compute lives in lib/ltcga-v175.js (fuzzed, zero non-finite leaks). PACSLAC is excluded by design (licensed/copyright-gated, fails the free-reproducibility bar).

Long-Term Care & Geriatric Assessment program: behavioral symptoms & observational delirium / mood screens (spec-v172, v174, +5 → 745)

spec-v174 is the program's second feature spec. The catalog already carried the interview-based delirium tiles (cam, cam-icu, 4at) and the self-report mood screens (gds15, phq9). The long-term-care floor needs the nurse-observation screens a charge nurse completes from a whole shift, and the dementia-specific behavioral instruments those self-report scales cannot cover. v174 ships all five proposed — each item value, per-item range, and band re-fetched and cross-verified against ≥ 2 independent sources (spec-v97):

Two findings of the cross-verification are worth recording. The interrai-abs draft per-item range (0–4, which would give a 0–16 total) was corrected to 0–3 / 0–12 against the CIHI job aid — the kind of off-by-one a re-fetch catches that recall does not. And cmai is reported as a frequency quantifier with no total severity band, because the CMAI manual explicitly advises against summing a severity score; the tile surfaces the three most-cited factor subscales (aggressive / physically non-aggressive / verbally agitated) and notes that factor membership varies by population. cornell-csdd surfaces inline that gds15 / phq9 self-report scales are not valid in moderate-to-severe dementia — the reason the Cornell scale exists. All five are Group G, Class A, trip no ISSUER_PATTERN, and carry boundary worked examples with band-flips (Nu-DESC 1→2, DOSS 2→3, Cornell 10→11 and 18→19, ABS mild/moderate and moderate/severe, CMAI floor/ceiling). The compute lives in lib/ltcga-v174.js (fuzzed, zero non-finite leaks).

Data-Sourced Reference-Table program: CDC growth charts, and a principled deferral (spec-v168, v169, +2 → 737)

spec-v168 is the fourth-pass program. The first three passes saturated the formula-shaped surface — closed-form scores a clinician could carry in their head. What remained were the instruments defined by a published reference table rather than an equation: pediatric BP percentiles, the transplant allocation indices, and the preterm/fetal growth standards. These were deferred on purpose by earlier passes, because shipping them safely requires the spec-v141 verbatim-fetch discipline — fetch the source bytes to disk, parse them programmatically, cross-verify against a second independent reproduction, and never hand-transcribe.

This pass shipped the two tiles whose source met that bar and deferred the other five that did not — the project's safety doctrine forbids shipping a clinically load-bearing calculator from a source it cannot fetch and verify.

The two shipped tiles are the percentile companions to the already-live peds-bmi-percentile (CDC BMI-for-age) and who-growth-zscore (WHO 0–2 yr), reusing the same interpLMS/lmsToZ infrastructure fuzzed since spec-v141.

Both are Class A with gate-forced citation-staleness rows (the "CDC" acronym trips the issuer pattern; the 2000 standard is fixed and does not drift). The deferrals are tracked, with their re-open condition, in docs/scope-data-sourced.md. The cross-verification is the point worth dwelling on: because the CDC files carry both the LMS coefficients and the printed percentiles, a single verbatim file is its own second source — the compute is correct iff it reproduces the publisher's own percentile columns, which it does to machine precision.

US-defaults & localization (spec-v184)

A clinician-perspective QA pass found the product was clinically strong but did not consistently present US defaults. spec-v184 fixed that as a presentation-only remediation (no tile added, no compute unit changed, every META.example and deep-link hash byte-identical):

• en-US locale declared consistently (<html lang>, JSON-LD inLanguage, every built page, the a11y assertion) to match og:locale.

• American-English copy across rendered bands, labels, banners, and category values (oedema→edema, haemoglobin→hemoglobin, tumour→tumor, ionised→ionized, noradrenaline→norepinephrine, …). A new lint gate, scripts/check-us-english.mjs, makes a British spelling in a user-facing string a build failure while leaving citations, journal abbreviations, and official instrument names (e.g. the mJOA "Japanese Orthopaedic Association") untouched.

• US date display — the Naegele due date renders MM/DD/YYYY (ISO) and the appeal-letter / HIPAA-request templates stamp Mon D, YYYY, via pure usDate/usDateLong formatters; the canonical ISO return value is unchanged.

• US-customary unit affordances — TEMP_UNITS (°C|°F) and HEIGHT_UNITS (cm|in) join the existing WEIGHT_UNITS (kg|lb); the canonical unit is always the default option, so unitNum still feeds the compute path metric and the documented examples reproduce exactly. The first wave shipped the toggles on the three energy tiles; the §4.3/§4.4 follow-on wave then closed the sweep so that every numeric temperature input now offers °F (NEWS2, MEWS, MEOWS, Truelove-Witts, the HScore, SNAPPE-II, APACHE II, SAPS II, and the CPIS) and every metric-only height input now offers inches (predicted spirometry, the IWPC and Gage warfarin dosers, bedside Schwartz eGFR, ARDSnet predicted body weight, pediatric BMI-for-age, the Ireton-Jones energy estimate, and GNRI). The weights entered alongside those converted heights (warfarin, peds BMI, Ireton-Jones, GNRI) gained the kg|lb toggle in the same change. The two height fields whose canonical unit is already inches stay inch-first — they already present the US default.

Cross-Discipline Completion program: EBM bedside math, ophthalmology, radiology classification, PK & one-formula gaps (spec-v162, v163–v167, +18 → 735)

spec-v162 is the third-pass program. The first two passes closed under-represented specialties (Post-Parity Coverage, v150) and subspecialty quantification (Subspecialty Depth, v157). This pass reaches the deterministic, free tools a clinician uses that a calculator catalog rarely indexes — the disciplines that live outside the usual "clinical score" framing. Five feature specs ship 18 tiles (a net +18 over the prior close), bringing the catalog to 735; every formula, point table, and threshold was re-fetched and cross-verified against ≥ 2 independent sources at implementation (the spec-v97 discipline), and all five modules are fuzz-covered by the spec-v59 harness with zero non-finite leaks.

Two things worth calling out. The program needed the one vocabulary edit the prior programs did not: clinical-epidemiology, ophthalmology, optometry, radiology, medical-physics, and audiology were added to the specialty-coverage.test.js closed vocabulary. And the program's actual delta is +18, not the nominal +19 — lithium-maintenance was deferred at implementation because the Cooper 1973 band table cannot be cross-verified to two independent sources (primary paywalled, secondary image-only, and the published equation does not cleanly reproduce the band table); it is parked with crib-ii / gail-bcrat. On citations, none of the 18 trips the check-citations issuer pattern (ACR, AAPM, ICRP, ASHA are not in it), so every tile is Class A with no staleness row. See docs/spec-v162.md.

Subspecialty Depth program: echocardiography, neuro/spine disability, rheumatology PRO & SLE, endocrine/metabolic math (spec-v157, v158–v161, +17 → 717)

spec-v157 is a second-pass program. Where the Post-Parity Coverage program (v150) closed the under-represented specialty gaps, this one closes the deeper subspecialty quantification gaps a finer read surfaces — most visibly that echocardiography, one of the most-performed studies in medicine, had a single quantification tile (aortic-valve-area). Four feature specs ship 17 tiles (700 → 717); every formula, weight, and partition was re-fetched and cross-verified against ≥ 2 independent sources at implementation (the spec-v97 discipline), and all four modules are fuzz-covered by the spec-v59 harness with zero non-finite leaks. The catalog ledger is docs/scope-subspecialty-depth.md.

Three things worth calling out. EDSS is the hardest scale to make deterministic — a precise Functional-System→step rating is not fully algorithmic, so the tile implements the standard simplified FS-count table for the low range and the authoritative ambulation anchors for ≥ 4.0, reporting the higher of the two (a wheelchair-dependent patient is never EDSS 2.0 because the FS table is low) and pointing the user to a trained Neurostatus rating for a definitive score. On citations, only calcium-phosphate-product (KDIGO) trips the check-citations issuer pattern and carries a documentation-only staleness row; ASE/EACVI, EULAR/ACR, the Endocrine Society, and ASPEN are all spelled out or out-of-pattern, so the other 16 tiles are Class A. And the program's actual delta is +17, not the draft's nominal +18 — the v157 draft carried a known running-count off-by-one; UTILITIES.length is the source of truth and the 13 catalog-truth surfaces agree at 717. See docs/spec-v157.md.

Rheumatology PRO & obstetric classification: BASDAI, BASFI, ESSDAI, Robson (spec-v156, +4 → 700) — Post-Parity Coverage program complete

spec-v156 is the sixth and closing feature spec of the Post-Parity Coverage program (spec-v150). v147/v148 shipped the physician-derived rheumatology activity scores (cdai-ra, sdai-ra, sledai-2k, asdas, ffs-2011); v156 completes the patient-reported axial-spondyloarthritis axis (basdai activity, basfi function), adds the standard Sjögren systemic-activity index (essdai), and ships the WHO-endorsed Robson Ten-Group cesarean-audit classifier alongside meows/bishop. The four computes live in lib/rheum-ob-v156.js + views/group-v156.js (RV156), fuzz-covered by the spec-v59 harness, every weight and decision rule re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline).

Three correctness anchors. BASDAI averages the two morning-stiffness items (Q5, Q6) before adding them — a unit test asserts the pair is not summed flat. ESSDAI's per-level printed value is already weight × level, so the total is a direct sum (max 123); the table preserves the structural quirks that secondary sources routinely flatten — constitutional/glandular/biological have no high level, and CNS has no low level. And Robson is asserted mutually-exclusive and total: a unit test enumerates all 144 input combinations and checks each maps to exactly one of the ten groups. On citations, essdai's EULAR and robson's WHO endorsement are documentation-only — EULAR is not in the check-citations issuer pattern (so no staleness row, matching the v147/v148 ACR/EULAR precedent), and Robson's WHO endorsement is kept out of the machine-read citation field. With v156 the Post-Parity Coverage program is complete (679 → 700). See docs/spec-v156.md.

Suite completions: MIPI, Forrest, Wagner DFU, University of Texas DFU (spec-v155, +4 → 696)

spec-v155 is the fifth feature spec of the Post-Parity Coverage program. Five suites were complete except for one well-known member each: the lymphoma-index suite (nccn-ipi, r-ipi, flipi) had no mantle-cell index, the upper-GI-bleed suite (gbs, rockall, aims65, oakland) had no endoscopic-stigmata anchor, and wifi graded limb threat but the diabetic-foot wound-grading systems were absent. The four tiles live in lib/suites-v155.js + views/group-v155.js (RV155), fuzz-covered by the spec-v59 harness, every coefficient and class re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline).

Two correctness anchors. MIPI's WBC is the absolute count per microliter inside the log — the Hoster erratum explicitly warns that inserting WBC as thousands/µL gives the wrong result (for 8000/µL use log₁₀(8000)=3.903, not log₁₀(8)), so the field is labelled "per µL, absolute" and a unit test locks the contract; the log₁₀ domain (LDH/ULN/WBC/age must be > 0) is the chief NaN path and returns a surfaced complete-the-fields fallback. And PRECISE-DAPT was deferred, not shipped: its published bleeding score is a restricted-cubic-spline continuous nomogram with no verbatim per-variable point table reproducible across ≥ 2 sources, so it is parked with crib-ii/gail-bcrat rather than approximated — the delta is +4, not the proposed +5. See docs/spec-v155.md.

Function, falls & palliative performance: Berg, TUG, Tinetti POMA, PPSv2 (spec-v154, +4 → 692)

spec-v154 is the fourth feature spec of the Post-Parity Coverage program. The catalog already carried fall-risk prediction (morse-falls, hendrich-ii) and frailty screens, but no performance-based mobility/balance measure, and palliative care had ecog-karnofsky but not the Palliative Performance Scale that anchors hospice eligibility. These four complete that axis: the two standard balance/gait batteries, the single most-used bedside mobility screen, and the hospice functional anchor. They live in lib/function-v154.js + views/group-v154.js (RV154), fuzz-covered by the spec-v59 harness, with every range, threshold, and band re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline).

Two correctness anchors. The Berg < 45 cutoff is strict — a score of exactly 45 sits on the lower-risk side, exercised by a 44/45 boundary test. And PPS is not a single dropdown: each column descriptor maps to a set of consistent levels (ambulation "Full" spans 100/90/80%), so the level is the best horizontal fit computed by intersecting the column candidate-sets left-to-right; a rightward column that conflicts with the leftward-established set is overridden by leftward precedence and flagged, never forced into an empty result. A unit test drives the read-leftward case where two columns disagree. See docs/spec-v154.md.

Urology & men's-health symptom scores: IPSS, IIEF-5/SHIM, OABSS (spec-v153, +3 → 688)

spec-v153 is the third feature spec of the Post-Parity Coverage program. The catalog already carried the urologic oncology math (psa-density, psa-velocity, psa-doubling-time, prostate-volume, gleason-grade-group, damico-prostate-risk, capra-score) and the stone scores, but none of the validated symptom-score instruments that drive benign-disease management — BPH/LUTS, erectile dysfunction, overactive bladder. These three are the standard, free, self-administered questionnaires. Each is a bounded item sum over fixed-range selects; an unanswered item surfaces a complete-the-fields fallback rather than an undercounted total. They live in lib/urology-v153.js + views/group-v153.js (RV153), Group G, Class A, fuzz-covered by the spec-v59 harness. Every item range, band cutoff, and gating rule was re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline).

Two correctness anchors. The IPSS quality-of-life item is never added into the 0–35 symptom total (a common scoring error) — a unit test asserts the total is invariant to the QoL value. And the OABSS surfaces its diagnostic gate: a high total driven by frequency alone (urgency item < 2) is flagged as not meeting the overactive-bladder symptom definition, rather than implying OAB from the total. See docs/spec-v153.md.

Nutrition & energy expenditure: predictive REE/BEE equations (spec-v152, +5 → 685)

spec-v152 is the second feature spec of the Post-Parity Coverage program. The catalog already had nutrition screening (must-nutrition, nrs2002, nutric, mnutric, refeeding-risk) and a weight-based icu-nutrition-target, but no predictive resting/total energy-expenditure equation — the number every dietitian starts from. These five fill that gap: the ambulatory standard, the classic comparator, the lean-mass equation, and the two ventilated-patient equations that approximate indirect calorimetry when a metabolic cart is unavailable. They live in lib/nutrition-energy-v152.js + views/group-v152.js (RV152), Class A, fuzz-covered by the spec-v59 harness. Every coefficient was re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline).

Two correctness anchors. Penn State is a three-way branch, not two — the modified (2010) constants apply only when BMI ≥ 30 and age ≥ 60; an obese patient under 60 still uses the standard 2003b form (a routing trap a unit test pins). And the 1997-revised Ireton-Jones constants (1784/244/239/804) are used, not the distinct 1992 set (1925/281/292/851) — the two were cross-verified side-by-side at implementation. See docs/spec-v152.md.

Dermatology severity: opening the Post-Parity Coverage program (spec-v151, +4 → 680)

spec-v150 charters the Post-Parity Coverage program — the successor to the completed spec-v100 MDCalc-Parity program — which asks one question, "have we included every calculator a healthcare worker would actually use?", and closes the under-represented-specialty gaps that the acute-care sweep left open. spec-v151 is its first feature spec: it fills the dermatology severity gap (the catalog had no scored-severity dermatology tile) with the four indices a dermatology clinic reaches for daily. All four live in lib/derm-v151.js + views/group-v151.js (RV151), Class A, fuzz-covered by the spec-v59 harness. Every region/item weight and band was re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline).

Two correctness anchors. EASI's region weights are age-dependent — children (< 8 yr) use head 0.2 / lower 0.3 where adults use head 0.1 / lower 0.4 (head and lower-limb weights swap; the same intensity inputs give a different total, which a unit test pins) — and the published six-band Leshem 2015 strata are used rather than the spec draft's unverified four-band cut-set (a spec-correction recorded in the source-governance notes, cross-verified against DermNet and the Hanifin 2022 practical guide). SCORAD's dryness item is graded on uninvolved skin (the classic trap), and the percentage → 0–6 area grade mapping is exercised by the spec-v59 fuzz harness. See docs/spec-v151.md.

Rheumatology, palliative & pharmacy: closing the spec-v100 program (spec-v148, +7 → 676, program complete)

spec-v148 is the closing feature spec of the entire spec-v100 MDCalc Parity Completion program. It adds the spondyloarthritis activity score, the vasculitis prognosis score, and the giant-cell-arteritis classification the rheumatology surface still lacked beside das28; the two free palliative-prognosis substitutes; an opioid equianalgesic rotation converter (distinct from the surveillance opioid-mme); and the Naranjo ADR causality scale. All seven live in lib/rheum-v148.js + views/group-v148.js (RV148), fuzz-covered by the spec-v59 harness. Every coefficient, weight, cutoff, and equianalgesic constant was re-fetched and cross-verified against ≥ 2 independent authoritative sources (the spec-v97 discipline).

Three correctness anchors. The ASDAS-CRP and ASDAS-ESR variants do not share their four NRS coefficients (0.12/0.06/0.11/0.07 vs 0.08/0.07/0.11/0.09) — a common transcription trap; both are computed and reported when both inputs are present, and a CRP < 2 mg/L is floored to 2 before the log term. The 2022 GCA biopsy and ultrasound-halo findings are one combined +5 item, not two, and the age-≥ 50 entry is enforced first. opioid-conversion is domain-guarded: the equianalgesic factors are fixed positive constants, a zero/blank/negative dose surfaces a complete-the-fields fallback rather than Infinity/NaN, and the high-stakes rotation carries the spec-v11 §5.3 second-check note. The proposed eighth tile valproate-correction was deferred — the spec's citation was wrong, the Hermida-Tutor free-fraction table could not be cross-verified to ≥ 2 sources, and a 2018 validation found it clinically inaccurate (parked with crib-ii / gail-bcrat / gwtg-hf / ROKS). With v148 the spec-v100 program is complete: 432 → 676. See docs/spec-v148.md.

MDCalc parity completion: the cardiology / vascular / lipid surface (spec-v100 program, Wave 1: spec-v101 → spec-v105, +25 → 457, complete; Wave 2 complete: spec-v106 → 463, spec-v107 → 467, spec-v108 → 473, spec-v109 → 478, spec-v110 → 483, spec-v111 → 487, +30 → 487; Wave 3 complete (Critical care & pulmonary): spec-v112 → 492, spec-v113 → 495, spec-v114 → 501, spec-v115 → 506, +19 → 506; Wave 4 underway (Neurology / neurosurgery / psychiatry): spec-v117 → 512, spec-v118 → 517, spec-v119 → 521, spec-v120 → 526, spec-v121 → 530, spec-v122 → 533, spec-v123 → 538 (Wave 4 complete), +32 → 538; Wave 5 complete (GI / hepatology / nephrology / acid-base / urology): spec-v124 → 544, spec-v125 → 549, spec-v126 → 555, spec-v127 → 559, spec-v128 → 564, spec-v129 → 570, spec-v130 → 576, spec-v131 → 584 (closes the wave at +5; ROKS deferred), +43 → 581 from the wave; live catalog 584 incl. the standalone spec-v149 EMS parity, +3; Wave 6 underway (Heme / onc / endocrine / ID): spec-v132 → 589 (thrombotic microangiopathy & coagulopathy, +5), spec-v133 → 593 (warfarin start-up — IWPC + Gage PGx dose, Kovacs 10 mg + Crowther 5 mg nomograms, +4, all four shipped after an adversarial re-fetch confirmed the Gage and Kovacs tables), spec-v134 → 599 (plasma-cell & myeloid-neoplasm staging — ISS, R-ISS, R2-ISS, Mayo MGUS, DIPSS, DIPSS-Plus, +6), spec-v135 → 604 (lymphoma / CLL prognostic indices — R-IPI, NCCN-IPI, GELF, Hasenclever IPS, CLL-IPI, +5), spec-v136 → 609 (endocrine / metabolic indices — HOMA-IR, QUICKI, TyG index, metabolic syndrome, OST/ORAI DXA pre-screen, +5), spec-v137 → 614 (infectious-disease scores — ISARIC 4C mortality, COVID-GRAM, Candida score, VACS index, RegiSCAR DRESS, +5, Wave 6 complete), +30 → 614; Wave 7 underway (Obstetrics / pediatrics / neonatal): spec-v138 → 620 (obstetrics & maternal-fetal medicine — Hadlock EFW, fullPIERS, miniPIERS, AFI, Barnhart hCG rise, IOM weight gain, +6, Wave 7 opens), spec-v139 → 626 (gynecology decision rules — Flamm VBAC, ROMA, RMI I/II/III, IOTA Simple Rules, Rotterdam PCOS, POP-Q staging, +6), spec-v140 → 631 (pediatric & neonatal severity — Kaiser EOS, SNAPPE-II, RDAI/Tal, Clinical Dehydration Scale, Koff bladder capacity, +5; CRIB-II deferred pending a second independent source for the Parry 2003 matrix), spec-v141 → 635 (pediatric growth / developmental-age — CDC 2000 BMI-for-age percentile, WHO 2006 growth z-score, Tanner mid-parental height, AAP corrected gestational age, +4 of 6: peds-weight-est skipped as a live spec-v149 duplicate, gail-bcrat deferred — its NCI BCRA incidence/competing-hazard tables ship only as binary .rda, not verbatim-fetchable to cross-verify per spec-v97), Wave 7 complete, +21 → 635; Wave 8 underway (Surgery / anesthesia / ortho / rheum / geriatrics / pharmacy): spec-v142 → 641 (surgical & anesthetic risk — POSSUM + P-POSSUM logistic morbidity/mortality, SORT 30-day-mortality logistic, Goldman 1977 cardiac index, Wilson difficult-airway sum, Sutton Surgical Risk Scale, +6, Wave 8 opens; SORT corrected to no-ASA-II-coefficient + mutually-exclusive age bands, Surgical Risk Scale range corrected to 3–14 from the draft's 3–17), spec-v143 → 646 (frailty & geriatric-oncology screening — mFI-5, mFI-11, FRAIL Scale, VES-13, CARG chemo-toxicity, +5; VES-13's 4-point all-or-nothing disability rule and 75–84 age band restored over two mis-printed online reproductions), spec-v144 → 652 (orthopedic fracture classification — Gustilo-Anderson open fracture I/II/IIIA–C, Garden femoral-neck I–IV, Danis-Weber ankle A/B/C, Schatzker tibial-plateau I–VI, Salter-Harris physeal I–V, Neer proximal-humerus one- to four-part, +6; the Gustilo Type III subtype keyed to coverage/perfusion not wound size, Weber dated to the 1972 monograph over a common 1966 mis-citation), spec-v145 → 657 (orthopedic risk & osteoarthritis — Frykman distal-radius I–VIII, Mirels impending-fracture 4–12, Kellgren-Lawrence OA grade 0–4, Pittsburgh knee rule, compartment delta pressure, +5; the Frykman even/odd ulnar-styloid axis, the Mirels ≥9 prophylactic-fixation flip, KL grade ≥2 as the definite-OA threshold, the entry-gated Pittsburgh mechanism, and the strict ΔP<30 mmHg fasciotomy threshold all cross-verified against ≥2 sources), spec-v146 → 662 (spinal tumor & trauma classification — SINS oncologic-instability 0–18, Revised Tokuhashi metastatic-spine prognosis 0–15, Tomita surgical-strategy 2–10, TLICS thoracolumbar triage 0–10, SLIC subaxial-cervical triage 0–10, +5; fills the spinal-scoring gap beside the existing brain/cerebrovascular neurosurgical cluster, TLICS/SLIC incomplete-cord > complete-cord and the SLIC +1 continuous-compression modifier cross-verified against ≥2 sources), spec-v147 → 669 (rheumatology activity & classification — CDAI lab-free RA activity 0–76, SDAI CRP-adding RA activity 0–86, 2010 ACR/EULAR RA classification 0–10 with the ≥6 definite flip, SLEDAI-2K SLE activity 0–105, 2015 ACR/EULAR gout with the MSU-crystal sufficient bypass and ≥8 threshold, CASPAR psoriatic-arthritis entry + ≥3 points, 2016 revised ACR fibromyalgia WPI/SSS dual-threshold, +7; the SDAI CRP mg/dL-vs-mg/L unit trap, the two negative gout items serum-urate <4 = −4 and MSU-negative synovial = −2, the SLEDAI-2K ongoing-activity credit, and the 2016 fibromyalgia somatic-count-not-severity simplification all cross-verified against ≥2 sources), spec-v148 → 676 (rheumatology / palliative / pharmacy, the program-closing spec — ASDAS spondyloarthritis activity, FFS-2011 vasculitis prognosis, 2022 ACR/EULAR giant-cell-arteritis classification, PPI and PaP palliative-prognosis, an opioid equianalgesic rotation converter, and the Naranjo ADR causality scale, +7; the proposed eighth tile valproate-correction was deferred — the spec's citation was wrong, the Hermida-Tutor free-fraction table could not be cross-verified to ≥2 sources, and a 2018 validation found it clinically inaccurate), +41 → 676, spec-v100 program complete)

With the spec-v85 program complete, spec-v100 charters the MDCalc Parity Completion program — a roadmap that closes the remaining gaps against the instruments a clinician expects to find — and opens Wave 1 (cardiology / electrophysiology / vascular / lipids). Wave 1 is now complete: five feature specs shipped (432 → 457, +25 — one below the projected +26 because spec-v102 deferred gwtg-hf), all under the same determinism / primary-source-citation / output-safety doctrine, each tile fuzz-covered by the spec-v59 harness:

• spec-v101 (+5 → 437) — atrial-fibrillation stroke-risk and QT instruments beside the existing combined chads view: chads2 (Gage 2001 + NRAF stroke-rate table), cha2ds2-va (2024 ESC, sex point removed), chads-65 (2020 CCS/CHRS pathway), atria-stroke (Singer 2013), tisdale-qtc (inpatient QT-prolongation risk). lib/cardio-v101.js.

• spec-v102 (+4 → 441) — heart-failure prognosis, HFpEF likelihood, and cardiogenic-shock mortality: maggic (Pocock 2013, with the age×EF / SBP×EF interactions and the 0–50 mortality lookup), h2fpef (Reddy 2018), hfa-peff (2019 ESC HFA, Class B), cardshock-score (Harjola 2015). lib/cardio-v102.js. The proposed gwtg-hf was deferred, not shipped — its per-band point table could not be verified from a reachable primary source, and this catalog does not ship fabricated scoring weights.

• spec-v103 (+6 → 447) — the CV-risk & prevention engines below, which complement, never replace, the existing ascvd (Pooled Cohort) and prevent tiles. Each cross-links them and states its derivation population so the clinician picks the right engine for the patient. lib/cvrisk-v103.js + views/group-v28.js.

• spec-v104 (+6 → 453) — ECG arrhythmia, aortic & syncope decision rules beside the existing ecg-axis / lvh-criteria tiles: the two wide-complex-tachycardia step algorithms (brugada-vt, vereckei-avr), the add-rs aortic-dissection pretest score with the optional D-dimer rule-out note, and the three ED syncope instruments (rose-syncope, egsys, oesil). All six Class A. lib/cardio-v104.js + views/group-v29.js. Detailed below.

• spec-v105 (+4 → 457, closes Wave 1) — peripheral-artery and cardiac-surgery risk: abi (the ankle-brachial index with the five PAD bands, Group E), the rutherford-fontaine PAD-stage mapper, the wifi SVS limb-threat classification, and the euroscore2 cardiac-surgery mortality engine. Adds the vascular-surgery specialty. lib/vascular-v105.js + views/group-v30.js. Detailed below.

Two correctness anchors. Coefficients were re-fetched, never recalled (the spec-v97 rule): the SCORE2 / SCORE2-OP region tables are a compiled constant transcribed from the published Eur Heart J 2021 supplement and cross-verified against the CRAN RiskScorescvd source — the implementation reproduces the two ESC published worked examples exactly (a 50-year-old smoker, SBP 140, TC 5.5, HDL 1.3 mmol/L lands at 5.9% in a low-risk region and 14.0% in a very-high-risk region for men; 4.2% and 13.7% for women), and the Framingham and Reynolds engines reproduce their papers' worked cases. Every logistic/Cox exponent is clamped to an overflow-safe range and every ln() term is domain-guarded, so the spec-v59 fuzz harness sees zero non-finite leaks; an unrecognized SCORE2 region returns a surfaced fallback rather than reading undefined coefficients, and an implausible negative remnant (LDL + HDL > TC) is flagged rather than printed. score2 / score2-op are Class B with citation-staleness rows (ESC region recalibration); the other four are Class A.

spec-v104 — ECG arrhythmia, aortic & syncope (+6 → 453)

The catalog had the ecg-axis and lvh-criteria ECG tiles but none of the bedside criteria an ED physician uses to call a wide-complex tachycardia, screen for aortic dissection, or risk-stratify syncope. spec-v104 fills that gap with six published, deterministic instruments — two boolean step algorithms, one category count, and three point/criterion scores:

Source-governs correction (the spec-v97 rule, applied to the prose this time). The spec-v104 draft described EGSYS with "syncope during effort or supine (3)" as one item and the two −1 terms as scored on absence. Verification against the primary Del Rosso 2008 paper and MDCalc found both wrong: effort (+3) and supine (+2) are separate items with distinct weights, and the precipitating- factors / autonomic-prodrome terms score −1 when present (they argue toward reflex syncope). The implementation follows the source — true range −2 to +12 (the positive weights sum to the universally-cited maximum of 12), not the draft's −2 to +10 — and the deviation is logged in docs/audits/v12/egsys.md. The boolean step rules return a defined verdict on a fully-negative input; egsys bounds its signed sum and oesil / add-rs index fixed lookups by a clamped total, so the spec-v59 fuzz harness sees zero non-finite leaks. All six citations name a journal

• authors (the ADD-RS title's "guideline-based" phrasing names no society acronym), so all six are Class A — no citation-staleness rows.

spec-v105 — vascular & cardiac surgery (+4 → 457, closes Wave 1)

The catalog had no peripheral-artery-disease bedside math and no cardiac-surgery mortality engine. spec-v105 adds the four instruments a vascular or cardiac surgeon reaches for, and closes Wave 1 of the spec-v100 program:

Two correctness anchors, both re-fetched, never recalled (the spec-v97 rule). (1) The EuroSCORE II age coefficient is 0.0285181 — the Nashef 2012 Eur J Cardiothorac Surg Table 6 multivariate value. The spec draft carried 0.0666354, which is the legacy logistic EuroSCORE I age coefficient; compiling it would over-estimate every prediction. The 30-coefficient block was transcribed verbatim and cross-verified against two independent sources; the model reproduces the published worked example exactly (a 70-year-old dialysis-dependent woman with insulin diabetes, COPD, NYHA III, CCS-4, poor LV, and a recent MI for isolated elective CABG gives y = −2.126358 → 10.66%). The on-dialysis coefficient is lower than CrCl ≤ 50 without dialysis — a published quirk of the model, reproduced rather than "corrected." (2) The WIfI grid is the amputation-risk table, not the separate revascularization-benefit table; all 64 cells were cross-verified across two reproductions. abi guards its brachial divisor for > 0 (a blank/zero brachial returns a surfaced fallback, never ankle/0) and reads the band off the rounded ratio so the shown index matches its band; euroscore2 clamps its logistic exponent to [−40, 40] so an extreme fuzzed input returns a probability in [0, 1] rather than Infinity. rutherford-fontaine and wifi are Class B with citation-staleness rows (SVS reporting/classification standards, on-publication cadence); abi and euroscore2 are Class A.

spec-v106 — VTE workup algorithms (+6 → 463, opens Wave 2)

Wave 2 of the spec-v100 program turns to emergency / triage surface. The catalog already carried the front-line VTE pretest tools (wells-pe, wells-dvt, perc, years-pe) and the prognostic pesi / spesi set, but six standard venous-thromboembolism workup instruments — each occupying a distinct decision point in the pathway — were absent. spec-v106 adds them, all in Group G:

Coefficients re-fetched, never recalled (the spec-v97 rule), each cross-verified across the primary paper + MDCalc / a clinical reference. The 4PEPS item weights (age < 50 −2 / 50–64 −1, chronic respiratory disease −1, HR < 80 −1, chest pain + dyspnea +1, male +2, estrogen +2, prior VTE +2, syncope +2, immobility +2, SpO₂ < 95% +3, calf pain/edema +3, PE most likely +5) and the four tier cutoffs were transcribed verbatim; the original Geneva ABG bands are encoded in both kPa and the rounded mmHg conversions (PaCO₂ < 36 mmHg / < 4.8 kPa +2, PaO₂ < 48.7 mmHg / < 6.5 kPa +4, …) because the Wicki paper reports SI units. peged and 4peps are strategy selectors: a missing tier or D-dimer renders a complete-the-fields fallback rather than a verdict from a missing value. constans-uedvt carries the signed −1 term and keys its band on the signed sum, not its absolute value. Every total is clamped to its published range and the band is read off the clamped value, so the spec-v59 fuzz harness sees zero non-finite leaks. All six citations name a journal + authors (NEJM, JAMA Cardiol, Eur Respir J, J Thromb Haemost, Arch Intern Med, Thromb Haemost) — none trips the issuer pattern, so all six are Class A with no citation-staleness rows. lib/vte-v106.js + views/group-v31.js.

spec-v107 — ED decision rules & resuscitation (+4 → 467)

Wave 2 continues with four standard emergency-department decision rules and resuscitation-risk scores that sit in the gaps between the chest-pain (heart, edacs), head-CT (pecarn-head, catch-head), and ICU-physiology (apache2, qsofa-sofa) clusters already in the catalog. spec-v107 adds them, all in Group G:

Coefficients re-fetched, never recalled (the spec-v97 rule), each cross-verified against the primary paper + MDCalc. The notable catch was GO-FAR: the "−15 to 11" figure quoted by some secondary sources is the per-variable point spread, not the total-score range (−15…+76), and the ≥ 24 "very low" band is reachable only because MDCalc treats the admission/comorbidity items as independent additive rows (no mutual-exclusivity enforcement) — shipping the category cut-points without reconciling that contradiction would have mis-banded high scores. hear and go-far require the numeric age and render a complete-the-fields fallback when it is missing; every total is clamped to its published range and the band is read off the clamped value, so the spec-v59 fuzz harness sees zero non-finite leaks. All four citations name a journal + authors (Eur J Emerg Med, NEJM, JAMA Intern Med, AJRCCM) — none trips the issuer pattern, so all four are Class A with no citation-staleness rows. go-far carries the explicit posture that the score informs, never decides, a goals-of-care discussion. lib/eddecision-v107.js + views/group-v32.js.

spec-v108 — Trauma severity scores & decision rules (+6 → 473)

Wave 2 continues with six standard trauma severity scores and decision rules. The catalog carried iss-rts (ISS + Revised Trauma Score) and abc-mtp (the ABC massive-transfusion rule), but the benchmark outcome model, the modern severity index, the two massive-transfusion probability tools, the pupil-adjusted GCS, and the chest-CT rule-out were all absent. spec-v108 adds them — triss and niss in Group E, the rest in Group G:

Coefficients re-fetched, never recalled (the spec-v97 rule), each cross-verified against the primary paper + MDCalc. The TRISS blunt/penetrating coefficient sets shipped are the MTOS-1995 revision values MDCalc serves (the literal 1987 paper published a smaller first set) — the citation names both. The TASH logistic sign is −4.9 + 0.3·TASH (rejecting the −0.3 transcription variants some secondary sources carry), reproducing the published ~50% anchor near a total of 16, and the additive max is 31 (MDCalc) rather than the abstract's 28. triss and tash-score clamp their logistic exponent to a finite range so a fuzz-extreme ISS or TASH total resolves to a finite probability rather than Infinity; niss clamps each AIS and applies the AIS-6 → 75 convention; gcs-pupils bounds the index to 1–15. All six citations name a journal + authors (J Trauma, World J Surg, J Neurosurg, PLoS Med) — none trips the issuer pattern, so all six are Class A with no citation-staleness rows. lib/trauma-v108.js + views/group-v33.js.

spec-v109 — Trauma classification & soft-tissue infection (+5 → 478)

Wave 2 continues with five standard trauma-classification and soft-tissue- infection decision rules. The catalog carried the trauma-physiology scores (iss-rts, triss, the massive-transfusion tools) and the skin-infection disposition surface, but the screening, grading, and salvage instruments below were absent. spec-v109 adds them, all in Group G:

Criteria re-fetched, never recalled (the spec-v97 rule), each cross-verified against the primary paper + MDCalc / RadioGraphics / EAST guideline. The notable transcription catches: the AAST 2018 contained-vs-extending vascular rule is grade-specific per organ (spleen contained IV / beyond V; liver and kidney contained III / beyond IV) — aast-organ-injury walks it as a per-organ decision tree (the anatomic-finding select rebuilds when the organ changes), not a browsable atlas, and returns the higher of the anatomic and vascular grade. The MESS ischemia-time doubling raises the ceiling from 11 to 14 and is applied before summing; mangled-extremity renders that it doubled. The LRINEC CRP threshold is 150 mg/L = 15 mg/dL (the most common unit-confusion error), and the probability bands (≤ 5 / 6–7 / ≥ 8) are distinct from the ≥ 6 suspicion cutoff. denver-bcvi and aast-organ-injury are Class B — the EAST/society screening criteria and the periodically-revised AAST scales each carry a docs/citation-staleness.md row (documentation-only: the citations name a journal

• authors, not an issuer acronym, so the row is not gate-forced but records the edition in force); the other three are Class A. None authors a CTA, debridement, antibiotic, or amputation order — the mangled-extremity posture note states explicitly that the score informs, never dictates, the salvage decision. lib/traumaclass-v109.js + views/group-v34.js.

spec-v110 — Toxicology dosing & dialysis decisions (+5 → 483)

Wave 2 continues with five standard poison-center / ED toxicology dosing and dialysis-decision tools. The catalog carried the acetaminophen-nomogram treatment line and the serotonin-toxicity / salicylate-toxicity / toxic-alcohol decision rules, but the high-frequency dosing math a clinician does by hand at the bedside was absent. spec-v110 adds it — the four dosing instruments in Group F (Medication & Infusion), the dialysis decision in Group G:

Every dosing tile renders the high-stakes second-check caveat (the spec-v100 §2 clause-5 requirement): the tile computes the figure, but the indication, timing, and route stay with the clinician, poison center, and local protocol. Formulas re-fetched, never recalled (the spec-v97 rule). Two source-governance catches: (1) the NAC 110-kg dosing-weight cap is a Math.min that clamps a 120-kg patient's bag doses to the 110-kg values (16500 / 5500 / 11000 mg on the three-bag regimen) and the band says so; (2) lithium-extrip follows the EXTRIP source over the spec prose — the spec draft put the "expected time to a level < 1.0 mmol/L exceeds 36 h" limb in the recommended set, but Decker 2015 places it (with level > 5.0 and confusion) in the suggested set, so the tile does too. lithium-extrip is Class B (a docs/citation-staleness.md row records the EXTRIP 2015 edition in force); the four dosing tiles are Class A. lib/tox-v110.js + views/group-v35.js.

spec-v111 — Environmental & wilderness medicine (+4 → 487, Wave 2 complete)

Wave 2 closes with four standard environmental / wilderness-medicine severity scores and classifications. The catalog carried the hypothermia-rewarm Swiss-staging tile, but the reference severity grading for four common exposures was reachable nowhere. All four home in Group I (EMS & Field), cross-linked from Clinical Scoring (Group G):

Three source-governance catches (the spec-v97 re-fetch rule): (1) lake-louise-ams enforces the headache-required gate — a total ≥ 3 without a headache does not diagnose AMS, and the 2018 revision dropped the sleep item (so the score is 0–12, not the legacy 0–15); (2) snakebite-severity follows the source over the spec prose — Dart 1996 validated the SSS as a continuous severity index and defines no fixed minimal/moderate/severe total-score cutoffs (the 0–3/4–7/≥8 bands circulated online belong to a different modified 7-system instrument), so the tile reports the continuous total and labels its descriptor as a relative reading of the 0–20 range; (3) cauchy-frostbite Grade-4 prognosis is "functional sequelae" per the NEJM 2022 reproduction of Cauchy's table, not the unverified "general/systemic" paraphrase. szpilman-drowning is Class B (a docs/citation-staleness.md row records the Szpilman 1997 edition in force); the other three are Class A. Each tile renders the spec-v50 §3 field-posture note: it grades severity and frames mortality / amputation risk to inform triage and transport, it does not author a descent, antivenom, debridement, or amputation order. lib/enviro-v111.js + views/group-v36.js. This closes Wave 2 of the spec-v100 program (457 → 487, +30).

spec-v117 — stroke imaging & thrombolysis prognosis (+6 → 512, opens Wave 4)

Wave 4 (Neurology / neurosurgery / psychiatry) opens with the six imaging-prognosis and thrombolysis-risk instruments the stroke team computes the moment the NCCT/CTA is read and tPA is on the table. The catalog had the stroke severity and disposition tools (nihss, abcd2, ich-score) but not the imaging-derived scores that gate reperfusion. Five home in Clinical Scoring & Risk (Group G); ich-volume-abc2 is Clinical Math & Conversions (Group E):

The four point-sum scores re-fetch the published point tables and outcome bands verbatim (the spec-v97 discipline), cross-verified across the derivation papers, MDCalc, and PMC reproductions. Where a source does not publish a per-score rate — DRAGON's middle range (4–7) and THRIVE's middle band (3–5) — the tile bands it intermediate and quotes only the robust published extremes rather than inventing a number (the project's no-fabrication rule). HAT's and SEDAN's symptomatic-ICH series are reproduced verbatim from the papers. ich-volume-abc2 guards each diameter (non-negative, finite) and its division; aspects clamps 0–10. Each tile reports the score/volume, not the order (spec-v11 §5.3): the thrombolysis, thrombectomy, surveillance, and surgical decisions stay with the stroke team and local protocol. Five are Class A; aspects is Class B (an imaging-read convention applied through evolving reperfusion guidelines → a documentation-only docs/citation-staleness.md row). lib/neuro-v117.js + views/group-v117.js. This opens Wave 4 of the spec-v100 program (506 → 512, +6).

spec-v127 — nephrology prognosis & AKI staging: KFRE, RIFLE, AKIN, ultrafiltration rate (+4 → 559, Wave 5)

Wave 5 turns to the nephrology prognosis and acute-injury-staging instruments beside egfr-suite, ckd-staging, ktv-urr, and kdigo-aki. kfre, rifle-aki, and akin-aki are in Clinical Scoring & Risk (Group G); ufr-dialysis is Clinical Math & Conversions (Group E):

All four re-fetch the coefficients / criteria verbatim (the spec-v97 discipline), cross-verified across ≥ 2 independent sources. The KFRE re-fetch resolved two real traps: the model uses the North American baseline survivals (4-variable S₀ = 0.9750 at 2 years / 0.9240 at 5 years — not the 0.9365 non-North-American value), and the ACR term is in mg/mmol, so a US spot UACR in mg/g is divided by 8.84 before the log (several published calculators mislabel this). The logistic is overflow-safe (1 − S₀^exp(lp) clamped to 0–1, never NaN for an extreme linear predictor); RIFLE's Failure acute-rise limb is strict > 0.5 while AKIN's stage-3 limb is ≥ 0.5 (the operators genuinely differ — kept distinct); and ufr-dialysis guards its weight/hours denominators. Each tile reports the probability / class / rate, not management (spec-v11 §5.3). All four are Class A (journal + author citations — KDIGO / ADQI / AKIN acronyms deliberately kept off the strings — no staleness row). lib/nephro-v127.js + views/group-v127.js. Catalog 555 → 559, +4.

spec-v149 — roughlogic.com EMS-group parity: pediatric weight estimate, PALS vitals, drug draw-up volume (+3 → 579, standalone)

A standalone spec (not part of the spec-v100 program, which reserves v101–v148). A cross-catalog audit of roughlogic.com's EMS group (27 tools) found 24 already covered; v149 ports the 3 genuinely-missing pre-hospital / field calculators, all in EMS & Field (Group I) with the field audience, re-grounded in their primary clinical sources:

peds-vitals is Class B (the "AHA" citation trips ISSUER_PATTERN, so it carries a docs/citation-staleness.md row); the other two are Class A. lib/ems-v149.js + views/group-v149.js. Catalog 576 → 579, +3.

spec-v139 — gynecology decision rules: Flamm VBAC, ROMA, RMI, IOTA Simple Rules, Rotterdam PCOS, POP-Q (+6 → 626, Wave 7)

v139 continues Wave 7 with the general-gynecology decision-rule cluster — the labor unit, the gyn-onc clinic, the reproductive-endocrine clinic, and the urogynecology exam room. All six land in Clinical Scoring & Risk (Group G) and report the score / index / verdict and the source's framing without authoring a counsel / refer / image / treat directive in Sophie's voice (spec-v11 §5.3):

Every coefficient block, point weight, and threshold was re-fetched from a primary source and cross-verified across ≥2 independent sources, never recalled (the spec-v97 discipline). The roma-ovarian logistic guards its ln(HE4) / ln(CA-125) domains for non-positive markers and clamps its exponent to [−40, 40], so ROMA% is never NaN or Infinity; all six flow through the spec-v59 fuzz harness with zero non-finite leaks. flamm-vbac, rmi-ovarian, iota-simple-rules, and popq-staging are Class A; roma-ovarian and rotterdam-pcos are Class B (their citations name a journal/authors and the ESHRE/ASRM group — none in the issuer-acronym set — so their citation-staleness rows are documentation-only, not gate-forced). The catalog vocab has no urogynecology term, so popq-staging is tagged obstetrics-gynecology (flagged in the spec + audit). lib/gyn-v139.js + views/group-v139.js (RV139). Catalog 620 → 626, +6.

spec-v138 — obstetrics & maternal-fetal medicine: Hadlock EFW, fullPIERS, miniPIERS, AFI, Barnhart hCG rise, IOM weight gain (+6 → 620, Wave 7 opens)

v138 opens Wave 7 (obstetrics / pediatrics / neonatal) by bringing the obstetrics and maternal-fetal-medicine cluster onto the page beside the dating and induction tiles (due-date, preg-dating, bishop, bpp). hadlock-efw, afi, barnhart-hcg, and iom-gwg read in Clinical Math & Conversions (Group E); fullpiers and minipiers in Clinical Scoring & Risk (Group G). Each reports the estimate / probability / range and the source's framing without authoring an image / deliver / transfer / counsel directive in Sophie's voice (spec-v11 §5.3):

Every coefficient block and threshold was re-fetched from a primary source and cross-verified across ≥2 independent sources, never recalled (the spec-v97 discipline). Three source-governance catches matter: (1) in fullpiers, SpO₂ has no main effect — it enters only through the platelet×SpO₂ interaction, and the −0.0271 coefficient belongs to creatinine (a common mis-recall assigns it to SpO₂); creatinine has no quadratic term. (2) minipiers takes the natural log of gestational age and systolic BP and codes dipstick proteinuria as three categorical indicators, with the 2+ level carrying the published negative weight −0.218 (non-monotonic, as published). (3) The optional iom-gwg current-gain comparison (spec §2.6) is deliberately not shipped — the IOM publishes no cumulative point target at an arbitrary gestational age, so a deterministic "on track" verdict would require fabricating a value the source does not state; for an underweight twin pregnancy the tile reports that no IOM recommendation exists rather than inventing one. The two logistic models clamp their exponent to [−40, 40] and the Hadlock log₁₀ is range-checked before 10^x, so every output is finite; all six flow through the spec-v59 fuzz harness with zero non-finite leaks. hadlock-efw, fullpiers, minipiers, and barnhart-hcg are Class A; afi and iom-gwg are Class B (ACOG-aligned revisable thresholds → gate-forced citation-staleness rows). lib/ob-v138.js + views/group-v138.js (RV138). Catalog 614 → 620, +6; Wave 7 opens.

spec-v137 — infectious-disease scores: ISARIC 4C, COVID-GRAM, Candida score, VACS index, RegiSCAR DRESS (+5 → 614, Wave 6 close)

v137 closes Wave 6 by bringing the infectious-disease risk-score cluster onto the page beside the community-acquired-pneumonia severity tools (curb-65, psi, smart-cop). All five land in Clinical Scoring & Risk (Group G) and report the score / probability and the source's framing without authoring an admit / start- antifungal / diagnose directive in Sophie's voice (spec-v11 §5.3):

Every point table, coefficient block, and threshold was re-fetched from a primary source and cross-verified across ≥2 independent sources, never recalled (the spec-v97 discipline). Four source-governance decisions follow the source over the spec draft where they diverged: (1) covid-gram invents no risk tiers — the paper deliberately defines none — and discloses that its betas are ln of the published odds ratios and its intercept derives from the paper's 1-significant- figure constant (OR 0.001), so the absolute probability is framed as approximate; the logistic exponent is clamped to [−40, 40] so an extreme fuzzed predictor returns a probability in [0, 1], never Infinity. (2) vacs-index quotes only the two published mortality anchors over a continuous calibration curve — no intermediate per-band percentage is fabricated (the gwtg-hf / ROKS precedent) — and its FIB-4 sub-computation guards the platelet and √ALT denominators with pos(). (3) isaric-4c-mortality applies the corrected Table 2 (urea < 7 mmol/L, CRP in mg/L) and exposes a urea/BUN unit selector (BUN mg/dL = urea mmol/L × 2.8). (4) candida-score renders the original > 2.5 cut-off as the integer ≥ 3 threshold. All five are Class A (journal + author citations — no ISSUER_PATTERN trip, no citation-staleness row) and flow through the spec-v59 fuzz harness with zero non-finite leaks. lib/id-v137.js + views/group-v137.js (RV137). Catalog 609 → 614, +5; Wave 6 complete (584 → 614, +30).

spec-v136 — endocrine / metabolic indices: HOMA-IR, QUICKI, TyG index, metabolic syndrome, OST/ORAI DXA pre-screen (+5 → 609, Wave 6)

v136 brings the endocrine / metabolic index cluster onto the page beside eag-a1c (the live A1c↔average-glucose converter). The three insulin-resistance surrogates land in Clinical Math & Conversions (Group E); the metabolic-syndrome rule and the bone pre-screen land in Clinical Scoring & Risk (Group G). Each reports the index or verdict and the source's framing without authoring a diagnose / start-drug / order-DXA directive in Sophie's voice (spec-v11 §5.3):

Every formula and threshold was re-fetched from a primary source and cross-verified across ≥2 independent sources, never recalled (the spec-v97 discipline). Four design points: (1) the IR indices guard their domains — homa-ir/quicki/tyg-index require glucose, insulin (and TG) >0 and surface a valid:false fallback rather than leaking log(0) = −∞, a divide-by-zero (QUICKI's denominator is 0 exactly when insulin×glucose = 1), or a NaN; the log/product math is explicitly fuzzed. (2) metabolic-syndrome honors the "or on treatment" override on each drug-modifiable criterion and applies the sex- and population-specific waist cut-point (US/ATP III M 102/F 88 cm; IDF Europid M 94/F 80; Asian M 90/F 80) — so every IDF-positive patient is also Harmonized-positive, but not vice-versa. (3) osteoporosis-prescreen truncates the OST index toward zero (Math.trunc, not Math.floor — the −3.6 → −3 worked example disambiguates) and encodes the ORAI point table exactly (age 45–54/55–64/65–74/≥75 = 0/5/9/15, weight ≥70/60–69/<60 = 0/3/9, no-estrogen +2), with the ≥9 referral threshold. (4) only metabolic-syndrome is Class B (a revisable consensus definition → a documentation-only docs/citation-staleness.md row); the other four are Class A (fixed formulas, journal + author citations — no ISSUER_PATTERN trip). All five flow through the spec-v59 fuzz harness with zero non-finite leaks. lib/endo-v136.js

• views/group-v136.js (RV136). Catalog 604 → 609, +5.

spec-v135 — lymphoma / CLL prognostic indices: R-IPI, NCCN-IPI, GELF, Hasenclever IPS, CLL-IPI (+5 → 604, Wave 6)

v135 brings the lymphoma and CLL prognostic-index cluster onto the page beside flipi (the live follicular index) and ipss-r-mds. All five land in Clinical Scoring & Risk (Group G) and quote the source's outcome framing without authoring a treat-versus-observe recommendation in Sophie's voice (spec-v11 §5.3):

Every weight and threshold was re-fetched from a primary source and cross-verified, never recalled (the spec-v97 discipline). Three design points: (1) nccn-ipi is banded, not a simple factor count — age and the LDH normalized ratio contribute up to 3 and 2 points respectively, so the maximum is 8, not 5; the exact band edges (age 60 → 1 / 75 → 2; LDH ratio 3 → 1 / >3 → 2) are pinned by boundary tests. (2) gelf-criteria is an any-one-positive flag that reports the criteria status — it never emits a "start chemotherapy" order; the treat-versus-watch decision stays with the clinician (it complements flipi, not replaces it). (3) cll-ipi's 4/2/2/1/1 weights make the high → very-high flip fall at 6 → 7, pinned by a dedicated boundary test. All five are Class A (journal + author citations — no staleness row); the "NCCN" in nccn-ipi's name is not an issuer acronym in the citation string (Zhou et al, Blood), so the check-citations.mjs ISSUER_PATTERN does not fire. All five flow through the spec-v59 fuzz harness with zero non-finite leaks and surface a complete-the-fields fallback rather than a partial group. lib/lymphoma-v135.js + views/group-v135.js (RV135). Catalog 599 → 604, +5.

spec-v134 — plasma-cell & myeloid-neoplasm staging: ISS, R-ISS, R2-ISS, Mayo MGUS, DIPSS, DIPSS-Plus (+6 → 599, Wave 6)

v134 brings the plasma-cell and myelofibrosis staging cluster onto the page beside ipss-r-mds (MDS prognosis) and flipi (lymphoma index). All six land in Clinical Scoring & Risk (Group G) and quote the source's survival framing without authoring a treatment recommendation in Sophie's voice (spec-v11 §5.3):

Every threshold and weight was re-fetched from a primary source and cross-verified, never recalled (the spec-v97 discipline). Two source-governance corrections to the spec draft: (1) R2-ISS totals 0–5, not "0–3.0" — the draft conflated the IV-stratum threshold (which opens at 3.0) with the score ceiling; the true maximum is ISS-III(1.5) + LDH(1.0) + del(17p)(1.0) + t(4;14)(1.0) + 1q21(0.5) = 5.0, pinned by a max-score test. (2) The R-ISS recomputes the ISS from the raw β2M + albumin rather than trusting a separately entered stage, so the ISS → R-ISS chain cannot desync. The only weighted-2 term in DIPSS is hemoglobin <10 g/dL (the common coding trap, guarded by a dedicated test). Five are Class A (journal + author citations — no staleness row); myeloma-r-iss is Class B (an IMWG working-group definition) and carries a docs/citation-staleness.md row (documentation-only — the spelled-out "International Myeloma Working Group" does not match the issuer acronym set). All six flow through the spec-v59 fuzz harness with zero non-finite leaks and surface a complete-the-fields fallback rather than a partial stage. lib/onc-v134.js + views/group-v134.js (RV134). Catalog 593 → 599, +6.

spec-v133 — warfarin start-up: IWPC + Gage pharmacogenetic dose, Kovacs 10 mg + Crowther 5 mg initiation nomograms (+4 → 593, Wave 6)

v133 brings oral-anticoagulant start-up onto the page beside heparin-nomogram — the catalog's only other "compute the next dose from inputs" tool. All four tiles land in Medication & Infusion (Group F) and carry the spec-v100 §2 clause-5 high-stakes second-check caveat in their rendered output:

Every coefficient and nomogram cell was re-fetched from a primary source and cross-verified, never recalled (the spec-v97 discipline). The IWPC block was extracted from the NEJM 2009 supplementary appendix S1e itself: √(weekly dose) = 5.6044 − 0.2614·decades + 0.0087·height + 0.0128·weight + VKORC1 + CYP2C9 + race + 1.1816·inducer − 0.5503·amiodarone, then squared for mg/week — the height coefficient is 0.0087 (the pharmacogenetic model) not 0.0118 (the clinical model, the classic cross-wire). The Gage 12-coefficient log-linear block (dose = exp(0.9751 + 0.4317·BSA − 0.00745·age − 0.2066·CYP2C9*2 − 0.4008·CYP2C9*3 − 0.3238·VKORC1 + 0.2029·INR − …)) was confirmed verbatim against a validation reprint and reconciled against the original Gage Table-3 percentages (VKORC1 −28%/allele = e^−0.3238 − 1, etc.); it uses the DuBois BSA the paper itself cites, and carries no CYP4F2 term (that was added later to the IWPC model by Sagreiya 2010, not to Gage). The Kovacs table encodes the full Figure-1 structure — the day-3 INR sets days 3 and 4 (which differ), and the day-5 INR selects one of four sub-tables by the day-3 band, with the 1.5–1.9 day-3 range correctly split at 1.6/1.7 (resolving the common reproduction disagreement); the 63.835/INR maintenance formula (Kovacs Blood 2007) and the AAFP day-5 maintenance table (Pengo 2001) are deliberately excluded as different instruments. The Crowther table preserves the day-5 low band at INR < 2.0 (not < 1.5 like days 3–4). All four are Class A (journal + author citations — no staleness row) and flow through the spec-v59 fuzz harness with zero non-finite leaks; the PGx models surface a valid:false fallback rather than squaring/exponentiating a degenerate input into a spurious dose. lib/warfarin-v133.js + views/group-v133.js (RV133). Catalog 589 → 593, +4.

Provenance note: this spec was authored as +4 but its draft assumed the Gage and Kovacs tables were unverifiable; an adversarial re-fetch (two independent reproductions per table) cleared both to publication fidelity, so all four shipped together — and corrected the draft's claim that "Gage adds CYP4F2."

spec-v132 — thrombotic microangiopathy & coagulopathy: PLASMIC, French TTP, JAAM DIC, IPSET-thrombosis, CISNE (+5 → 589, Wave 6 open)

v132 opens Wave 6 (Heme / onc / endocrine / ID) of the spec-v100 program with the thrombotic-microangiopathy / coagulopathy cluster that sat conceptually beside the existing four-ts (HIT probability) and khorana (cancer-VTE) tiles but was reachable nowhere. All five home in Clinical Scoring & Risk (Group G):

Every point table was re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline), and three source-governance calls were made: (1) PLASMIC's active-cancer and transplant points score for the absence of the condition — the classic coding inversion, locked by a dedicated test; (2) the French TTP creatinine threshold ships inclusive (≤ 2.26 mg/dL / ≤ 200 µmol/L) per Coppo 2010, governing over the spec draft's strict < (a creatinine of exactly 2.26 scores the point); (3) jaam-dic is the 2006 revised criteria — the fibrinogen term was removed and the maximum is 8, not the older max-10 fibrinogen form that several secondary calculators still show.

A latent-bug note worth recording: the graded selects (ECOG, mucositis) revealed that Number(null) and Number('') both coerce to 0, which would have let a blank ECOG silently score as ECOG 0. The numeric coercion helper now rejects null / undefined / '' / booleans up front, so a blank graded field surfaces the complete-the-fields fallback instead of a false zero.

All five are Class A (journal + author citations — no staleness row), flow through the spec-v59 fuzz harness with zero non-finite leaks, render the spec-v50 §3 posture note, and treat a blank component as not-assessed rather than silently scoring 0. lib/heme-v132.js + views/group-v132.js (RV132). Catalog 584 → 589, +5.

spec-v131 — urology renal mass / kidney stone / torsion: CAPRA, R.E.N.A.L., PADUA renal, S.T.O.N.E., TWIST (+5 → 584, Wave 5 close)

v131 closes Wave 5 by completing the urology cluster v130 opened: the renal-mass anatomic-complexity scores, the PCNL stone-complexity score, the stone-cancer recurrence score, and the point-of-care testicular-torsion rule. All five home in Clinical Scoring & Risk (Group G):

Every point table was re-fetched and cross-verified against ≥ 2 independent sources (the spec-v97 discipline, dispatched as three parallel research passes), and several source-governance calls were made: (1) padua-renal is a collision rename — the catalog already has the unrelated VTE Padua Prediction Score (padua), so the renal score ships under a distinct id and the two are never cross-linked; (2) stone-nephrolithometry ships the original Okhunov PCNL area version (stone area = length × width mm², total 5–13), explicitly not the later ureteroscopy adaptation that scores size by diameter and runs 5–15; (3) the R.E.N.A.L. (A)nterior/posterior face is a non-scoring suffix (a/p/x), with an appended h for hilar tumours, not a points contribution; (4) the CAPRA Gleason axis is not the summed 2–10 score — a primary pattern 4/5 scores +3, a secondary-only 4/5 scores +1, and there is no +2 level (it jumps 1 → 3).

Design decision — ROKS deferred (a refusal, on purpose). The spec scoped a sixth tile, the Rule 2014 ROKS recurrence-of-kidney-stone nomogram. Its 2-/5-/10- year probability formula is published, but the per-variable points that feed it live only in a graphical nomogram; the papers publish hazard ratios, not a numeric point table or the points-scaling constant. Transcribing them would mean measuring pixel positions off a figure — fabrication. This program already refused exactly that once (gwtg-hf, spec-v102), so ROKS is deferred and its id reserved until an institutional coefficient appendix is available. v131 therefore ships +5, not +6. Refusing to ship a guessed clinical coefficient is a feature of the maintenance contract, not a gap.

All five are Class A (fixed published point tables; journal + author citations — no staleness row), flow through the spec-v59 fuzz harness with zero non-finite leaks, render the spec-v50 §3 posture note, and treat a blank component as not-assessed (a surfaced fallback) rather than silently scoring 0. lib/uro-v131.js + views/group-v131.js. Catalog 579 → 584, +5.

spec-v130 — urology prostate metrics & risk: prostate volume, PSA density/velocity/doubling-time, D'Amico, Gleason Grade Group (+6 → 576, Wave 5)

Wave 5 opens the urology surface. v130 adds the prostate-volumetry and PSA-kinetics math a urologist runs at the bedside plus the two canonical prostate-cancer-risk classifications. Four home in Clinical Math & Conversions (Group E) and two in Clinical Scoring & Risk (Group G):

All six re-fetch the formulas/coefficients verbatim (the spec-v97 discipline, dispatched as an independent research pass), cross-verified across ≥ 2 sources, and four source-governance calls were made: (1) prostate-volume — the ellipsoid coefficient is fixed at 0.52 (π/6 rounded to the dominant clinical/MDCalc convention; the alternate exact 0.5236 differs by ~0.7%), stated to the user. (2) damico-prostate-risk — the PSA boundary is strict (> 10), so a PSA of exactly 10 is Low, and the high-risk T-stage cut is T2c per the original paper, with the worst single feature governing the group. (3) psa-doubling-time — the doubling time is undefined for a stable or falling PSA, so the tile detects a non-rising PSA and returns "not rising" rather than a NaN / negative time. (4) psa-velocity — the validated method averages consecutive yearly rates over ≥ 3 measurements spanning ≥ 18 months, so the two-point rate is labeled the bedside approximation. The PSA-kinetics tiles report signed results; every denominator is guarded. Each tile reports the urologic quantity or risk class, not management (spec-v11 §5.3). All six are Class A (journal + author citations — no staleness row). lib/uro-v130.js + views/group-v130.js. Catalog 570 → 576, +6.

spec-v129 — acid-base compensation & gaps: Stewart SID/SIG, base excess, the three compensation formulas, urine osmolal gap (+6 → 570, Wave 5)

Wave 5 closes the acid-base surface: the catalog already had the anion gap (anion-gap-dd) and Winter's formula (winters), and v129 completes the compensation set, adds the physicochemical (Stewart) view, the hemoglobin-corrected base excess, and the urine osmolal gap. All six home in Clinical Math & Conversions (Group E):

All six re-fetch the formulas/coefficients verbatim (the spec-v97 discipline, dispatched as two independent research passes), cross-verified across ≥ 2 sources, and four source-governance calls were made: (1) stewart-sid-sig — the published Figge SIDe is pH-dependent, but the spec input set omits pH, so the weak-acid charges are fixed at the physiologic pH 7.4 (yielding albumin 2.8 mEq/L per g/dL and phosphate 0.59 mEq/L per mg/dL, both derived from the Figge coefficients); the assumption is stated to the user. (2) base-excess — the Van Slyke constants are kept as one matched NCCLS C12-T2 pair (24.8 ↔ 9.5/1.63, factor 0.0143; Lang & Zander 2002 warn against crossing editions), reproducing the published −13.0 mEq/L worked case. (3) the three compensation formulas use an explicit acute-vs-chronic selector (never inferred) and clamp every prediction to a physiologic range, comparing expected vs measured to flag a superimposed disorder. (4) urine-osmolal-gap uses the standard US-unit calculated-osmolality identity (urea-N ÷ 2.8, glucose ÷ 18). The SIG / base-excess / urine-gap tiles report signed results; every denominator is guarded. Each tile reports the acid-base quantity or the expected-vs-measured comparison, not management (spec-v11 §5.3). All six are Class A (journal + author citations — no staleness row). lib/acidbase-v129.js

• views/group-v129.js. Catalog 564 → 570, +6.

spec-v128 — renal excretion & dialysis math: FE-phosphate, FE-magnesium, nPCR/nPNA, standard Kt/V, electrolyte-free water clearance (+5 → 564, Wave 5)

Wave 5 continues into the renal-excretion and dialysis-math surface beside fena-feurea and ktv-urr. All five home in Clinical Math & Conversions (Group E):

All five re-fetch the formulas verbatim (the spec-v97 discipline), cross-verified across ≥ 2 independent sources, and three source-governance calls were made over the spec draft: (1) efwc's sign was inverted in the spec prose — the cross-verified convention (Rose 1986, the Frontiers 2018 review, ScienceDirect) is that a positive EFWC is net free-water excretion (raises plasma Na, toward hypernatremia) and a negative EFWC is retention (lowers Na, drives hyponatremia); the tile implements the corrected sign. (2) npcr-pna ships the two-point intradialytic-rise form (reproducing the published 1.24 g/kg/day worked example) and deliberately does not ship the Kt/V-coefficient form, whose first-of-week and last-of-week coefficient triplets are unrecoverable from open sources (no-fabrication, cf. the deferred gwtg-hf). (3) femg keeps the 0.7 free-fraction correction on the denominator (a minority calculator variant drops it and inflates FE by ~1/0.7). The fractional- excretion and kinetic tiles guard every denominator; efwc requires a nonzero plasma sodium and reports the signed result. Each tile reports the excretion fraction / adequacy quantity / signed clearance, not management (spec-v11 §5.3). All five are Class A (journal + author citations; KDOQI / FHN never reach ISSUER_PATTERN — no staleness row). lib/renal-v128.js + views/group-v128.js. Catalog 559 → 564, +5.

spec-v126 — GI disease activity & pancreatitis severity: CDAI, UCEIS, SES-CD, HAPS, Balthazar CTSI, modified Marshall (+6 → 555, Wave 5)

Wave 5 continues with the clinical-trial-standard IBD activity indices and the standard pancreatitis imaging / organ-failure scores. All six home in Clinical Scoring & Risk (Group G):

All six re-fetch the published weights / scales / thresholds verbatim (the spec-v97 discipline), cross-verified across ≥ 2 independent sources, resolving three genuine literature conflicts: UCEIS ships the 0-based 0–8 scale (the original 2012 paper was 1-based 3–11, later rebased); SES-CD's true maximum is 56 (the stenosis sub-total is capped at 11 because a non-passable stenosis ends the exam — not the naive 60 that even ECCO's widget shows); and HAPS uses strict < thresholds (the cutoff value itself is abnormal). cdai-crohns guards the standard-weight divisor and the log-free hematocrit/weight terms; modified-marshall guards the PaO₂/FiO₂ denominator and reports a blank system as not-assessed (the v93 glasgow-imrie pattern). Five are Class A; modified-marshall is Class B (the revisable Revised-Atlanta organ-failure definition → a documentation-only docs/citation-staleness.md row, on-publication cadence — its citation names the working group, not an issuer acronym, so it is not gate-forced). lib/gi-v126.js + views/group-v126.js. Catalog 549 → 555, +6.

spec-v125 — hepatology severity & encephalopathy: PELD, CLIF-C ACLF, GAHS, West Haven, HSI (+5 → 549, Wave 5)

Wave 5 continues with the severity and complication instruments hepatologists reach for in acute deterioration. All five home in Clinical Scoring & Risk (Group G):

All five re-fetch the published coefficients / bands / criteria verbatim (the spec-v97 discipline), cross-verified across ≥ 2 independent sources, with three governance points: GAHS uses SI units — blood urea in mmol/L and bilirubin in µmol/L (the one real GAHS pitfall; not BUN/mg/dL), so the tile takes SI natively; CLIF-OF circulation scores 3 for vasopressor use (the canonical Jalan 2014 definition, not the MAP < 65 of CLIF-SOFA that crept into one secondary table); and PELD uses the raw McDiarmid form (no ×10 — that is the UNOS allocation presentation), with labs floored at 1.0. The log tiles (PELD, CLIF-C ACLF, HSI) domain-guard every ln/ratio (blank/non-positive → complete-the-fields fallback, never ln(0) or divide-by-zero). Each tile reports the score/grade, not management (spec-v11 §5.3). All five are Class A (journal + author citations, no ISSUER_PATTERN trip → no staleness row). lib/hep-v125.js + views/group-v125.js. Catalog 544 → 549, +5.

spec-v124 — hepatology function & fibrosis: ALBI, MELD-XI, Forns, BARD, FLI, Lok (+6 → 544, opens Wave 5)

Wave 5 (GI / hepatology / nephrology / acid-base / urology) opens with the family of objective liver-function grades and serum fibrosis/steatosis surrogates a hepatologist reads beside the existing meld-childpugh and fib4 tiles. albi-grade and bard-score are in Clinical Scoring & Risk (Group G); the rest are Clinical Math & Conversions (Group E):

All six re-fetch the published coefficients verbatim (the spec-v97 discipline), cross-verified across ≥ 2 independent sources — and the re-fetch caught a real spec error: the spec draft labeled the Forns cholesterol input "mmol/L," but the −0.014 coefficient is calibrated to mg/dL (feeding mmol/L would make the term ~38× too small and grossly inflate the score), so the tile takes mg/dL. Other governance: ALBI uses the primary-paper −0.085 (not −0.0852); MELD-XI floors both labs at 1.0 before the log (the standard-MELD convention, documented — no rescaling, no creatinine cap); the logistic tiles (FLI, Lok) use an overflow-safe 1/(1+e^−x) so extreme inputs return 0 or 100/1, never Infinity; and every ln/log₁₀ argument is domain-guarded so a blank or non-positive value surfaces a complete-the-fields fallback rather than ln(0). Each tile reports the grade / score / probability, not management (spec-v11 §5.3). All six are Class A (journal + author citations, no ISSUER_PATTERN trip → no staleness row). lib/hep-v124.js + views/group-v124.js. Catalog 538 → 544, +6 — opens Wave 5.

spec-v123 — psychiatry public-domain instruments: AIMS, Bush-Francis, Barnes, SCOFF, CES-D (+5 → 538, closes Wave 4)

The Wave 4 closer adds five confirmed public-domain / free-to-use psychiatry instruments — the movement-side-effect, catatonia, eating-disorder, and depression scales whose license status permits an interactive build. (The copyrighted ones — BDI, PANSS, MoCA, EAT-26, … — stay on the spec-v100 §8 permanent-exclusion list.) All five home in Clinical Scoring & Risk (Group G):

All five re-fetch the item lists / scoring keys verbatim (the spec-v97 discipline), cross-verified across ≥ 2 independent sources, with three corrections worth noting: the BFCRS item order (Immobility/stupor is item 1, Excitement item 14 — not the reverse) and its six 0/3-binary items (12, 17–21); the CES-D reverse-scored items are 4, 8, 12, 16 (applied in-compute, so the all-"rarely" default correctly scores 12, not 0); and the CES-D ≥ 16 adult cutoff (not the child CES-DC's ≥ 15, a documented conflation deliberately avoided). Each tile is a screen or severity scale, not a diagnosis (spec-v11 §5.3): the assessment and treatment decision stay with the clinician. All five are Class A (journal/manual + author citations, no ISSUER_PATTERN trip → no staleness row), and each instrument's public-domain / free-to-use status is re-confirmed per spec-v100 §8 in docs/clinical-citations.md. lib/psych-v123.js + views/group-v123.js. Catalog 533 → 538, +5 — closing spec-v100 Wave 4 (Neurology / neurosurgery / psychiatry), which grew from 506 to 538 (+32) across v117–v123.

spec-v122 — general neurology & rehab: dementia type, spasticity, brainstem encephalitis (+3 → 533, Wave 4)

Wave 4 widens from the neuromuscular emergencies into three general-neurology and rehabilitation instruments that cross specialty lines. All three home in Clinical Scoring & Risk (Group G):

All three re-fetch the published weights / ordinal wording / criteria verbatim (the spec-v97 discipline), cross-verified across ≥ 2 independent sources. Three source-governance choices: (1) the hachinski weights were re-fetched, not recalled — stepwise deterioration is 1 point, not 2 (a common mis-recall the official ARIC/NIH form and every reproduction contradict; max 18, not the inflated value the wrong assumption gives); (2) modified-ashworth renders "1+" as a distinct ordinal step via string keys, never averaged or summed into a fractional grade; (3) bickerstaff does not gate the verdict on the anti-GQ1b antibody — it, the MRI lesion, and the CSF dissociation are supportive only (seronegative cases are recognized), and the tile frames the determination as a research/classification reading, not a validated gold standard, naming the GQ1b spectrum link to Miller Fisher syndrome and GBS. Each tile reports the score / grade / determination, not the order (spec-v11 §5.3): the diagnosis and management decision stay with the clinician and local protocol. All three are Class A (journal+author citations, no ISSUER_PATTERN trip → no staleness row). lib/neuro-v122.js + views/group-v122.js. Catalog 530 → 533, +3.

spec-v121 — neuromuscular emergencies: GBS & myasthenia (+4 → 530, Wave 4)

Wave 4 moves from epilepsy/headache/vertigo into the neuromuscular-emergency instruments a neurology and neurocritical-care service uses to predict respiratory failure and grade disease. The catalog had no GBS or myasthenia tools. All four home in Clinical Scoring & Risk (Group G):

All four re-fetch the published point weights / case definition verbatim (the spec-v97 discipline), cross-verified across the derivation papers and open-access reproductions (the PMC "Ten Steps" GBS review Box 3, the Bangladesh & Frontiers mEGOS validations, the Fokke 2014 Brain Brighton-table reprint, and the official MGFA Foundation classification PDF). Two source-governance choices, consistent with the project no-fabrication doctrine: (1) EGRIS publishes only the three banded category rates (4% / 24% / 65%) over a continuous logistic curve with no per-integer-score table, so the tile quotes the category rates; (2) mEGOS's per-score probability of inability to walk is published only as figure curves whose coefficients are not reported and which diverge by region, so the tile reports the total and a relative reading of the published range (the v111 snakebite-severity relative-range pattern), inventing no per-score percentage. Each tile reports the score / level / class, not the order (spec-v11 §5.3): the IVIG / PLEX / intubation / monitoring decision stays with the clinician and local protocol. All four are Class A (journal+author citations, no ISSUER_PATTERN trip → no staleness row). lib/neuro-v121.js + views/group-v121.js. Catalog 526 → 530, +4.

spec-v120 — epilepsy, headache & vertigo (+5 → 526, Wave 4)

Wave 4 continues from the stroke surface into the epilepsy-prognosis, headache-likelihood, and vertigo-localization rules a neurologist or ED clinician runs daily. The catalog had stroke scales and the dementia screens but not these five. All home in Clinical Scoring & Risk (Group G):

All five re-fetch the published point weights / lookup verbatim (the spec-v97 discipline), cross-verified across the derivation papers and the JAMA / MDCalc / PMC reproductions. Three source-governance choices, consistent with the project no-fabrication doctrine: (1) helps2b is ML-derived but ships as a compiled integer→risk lookup constant — no model runs at render time (spec-v100 §11), and the paper's collapsed ">95%" stratum folds scores 6 and 7 together rather than inventing a 7-specific figure; (2) stess has no published per-score mortality table, so the tile frames the favorable (0–2) / unfavorable (≥ 3) dichotomy and the high negative predictive value (~0.97), inventing no per-band percentage; and (3) mess-first-seizure's per-year treated/deferred recurrence grid is paywalled (Lancet Neurol Table 4), so the tile reports the confirmable risk-group ranges over a 3–5 year window rather than fabricating discrete annual cells. The id is distinct from the v109 mangled-extremity MESS (spec-v100 §4 collision audit). Each tile reports the score / classification, not the order (spec-v11 §5.3): the treat / admit / monitor / image decision stays with the clinician and local protocol. All five are Class A (journal+author citations, no ISSUER_PATTERN trip → no staleness row). lib/neuro-v120.js + views/group-v120.js. Catalog 521 → 526, +5.

spec-v119 — prehospital LVO triage & cerebrovascular diagnosis (+4 → 521, Wave 4)

Wave 4 continues from the in-hospital hemorrhagic grading out to the field. The catalog had the in-hospital stroke scales (nihss, ich-score) but not the prehospital LVO-triage tools the EMS crew runs, nor two cerebrovascular-diagnosis rules. All four home in Clinical Scoring & Risk (Group G):

All four re-fetch the published point weights / diagnostic logic verbatim (the spec-v97 discipline), cross-verified across the derivation papers and MDCalc / PMC / validation-cohort reproductions. Two source-governance catches resolved at implementation: FAST-ED totals 0–9, not the "0–10" MDCalc's UI labels (the item maxima — facial palsy caps at 1 — sum to 9; the "0–10" is a sum-of-fives artifact); and the CVT coma item is +2, not the stray "+5" that circulates (it reflects the published hazard ranking, where malignancy, coma, and deep-CVT carry the three highest hazard ratios → 2 points each). The Boston v2.0 logic is verbatim-confirmed across two independent reproductions of the Lancet Neurology source; boston-caa reports "Criteria not met" rather than inventing a category when a deep lesion, a missing age/presentation, or no qualifying marker is present. Each tile reports the score/category, not the order (spec-v11 §5.3): the destination, bypass, anticoagulation, and treatment decisions stay with the EMS crew, stroke team, and local protocol. cpsss/fast-ed/cvt-risk are Class A (journal+author citations, no ISSUER_PATTERN trip → no staleness row); boston-caa is Class B (a revisable consensus diagnostic definition → a documentation-only docs/citation-staleness.md row). lib/neuro-v119.js + views/group-v119.js. Catalog 517 → 521, +4.

spec-v118 — hemorrhagic stroke, SAH, IVH & aneurysm (+5 → 517, Wave 4)

Wave 4 continues on the hemorrhagic side the neuro-ICU and neurosurgery teams grade. v117 covered the ischemic-stroke imaging-prognosis scores; v118 adds the SAH-blood, IVH-burden, hematoma-expansion, and unruptured-aneurysm instruments. All five home in Clinical Scoring & Risk (Group G):

All five re-fetch the published point tables verbatim (the spec-v97 discipline), cross-verified across the derivation papers and PMC / validation- cohort reproductions. Two source-governance catches resolved at implementation: the modified Graeb maximum of 32 is reached only because the +1 expansion bonus is an independent additive modifier on each of the eight compartments (a naive "fill grade alone" reading sums to 24); and ELAPSS scores no earlier SAH as +1 (a prior treated bleed associates with lower growth risk of the remaining aneurysm). Each tile reports the grade/score, not the order (spec-v11 §5.3): the coiling, clipping, surveillance, and surgical decisions stay with the neurosurgery / neurocritical-care team. All Class A (journal+author citations, no ISSUER_PATTERN trip → no staleness row). lib/neuro-v118.js + views/group-v118.js. Catalog 512 → 517, +5.

spec-v115 — pulmonary nodule, PH & pleural infection (+5 → 506, closes Wave 3)

Wave 3 closes with five pulmonary decision rules a pulmonologist reaches for routinely. The catalog had the chronic-airways and acute-PE tools but lacked the incidental/screen-detected nodule malignancy models, the nodule-surveillance matrix, the PAH prognosis score, and the pleural-infection mortality score. All five home in Clinical Scoring & Risk (Group G):

Both nodule logistics re-fetch the published coefficients verbatim (the spec-v97 discipline) — the seven Mayo coefficients confirmed against the spec draft, and the full Brock equation cross-verified across the original NEJM appendix, Radiopaedia, and MDCalc, including the (size/10)^−0.5 − 1.58113883 power transform (the centering constant is 0.4^−0.5) and the age/count centering at 62 and 4. Both clamp the logistic exponent to [−40, 40] so a fuzzed extreme (a 1e9 mm diameter) yields 100%, never NaN/Infinity; Brock domain-guards size > 0. Each tile reports the probability/interval/score, not the order (spec-v11 §5.3): the surveillance, PET, biopsy, escalate-care, or drainage decision stays with the clinician and local protocol. Four are Class A (fixed coefficients/weights, journal+author citations — no staleness row); fleischner-2017 is Class B (revisable Fleischner Society 2017 guidance → a documentation-only docs/citation-staleness.md row). lib/pulmnod-v115.js + views/group-v40.js. This closes Wave 3 of the spec-v100 program (487 → 506, +19).

spec-v114 — COPD/bronchiectasis exacerbation & sleep (+6 → 501, Wave 3)

Wave 3 continues with six pulmonary and sleep-medicine decision rules. The catalog had the chronic-airways staging tools (gold-spirometry, bode-index, predicted-spirometry) and a sleep screen (stop-bang), but lacked the standard acute-COPD-exacerbation prognosis, bronchiectasis-severity, and sleep-disordered-breathing classifiers. All six home in Clinical Scoring & Risk (Group G):

The research re-fetch corrected three spec-draft errors (SOURCE governs): FACED Extension scores at ≥ 3 lobes (not ≥ 2) and Dyspnea at mMRC ≥ 3 (not ≥ 2); the BSI admission window is the prior 2 years and uses the MRC 1–5 scale, not mMRC. BAP-65 is a class derived from a count, not a 0–4 point sum — an 80-year-old with no acute variable is class II (0.9% mortality), not high-risk; the mechanical-ventilation rates for the lower classes rest on a single source, so the tile reports them qualitatively rather than publishing uncertain figures (spec-v97 no-fabrication discipline). Each tile reports the score/class/band, not the order (spec-v11 §5.3): the admit / ventilate / refer-for-sleep-study decision stays with the clinician and local protocol. Five are Class A (fixed derivation-paper weights, journal+author citations — no staleness row); ahi-odi-severity is Class B (revisable AASM criteria → a documentation-only docs/citation-staleness.md row). lib/pulm-v114.js + views/group-v39.js. Wave 3 reaches 501 (495 → 501, +6).

spec-v113 — dynamic fluid-responsiveness indices (+3 → 495, Wave 3)

Wave 3 continues with the three dynamic preload-responsiveness indices an intensivist uses at the bedside to decide whether a fluid bolus will help. The catalog computed static hemodynamics (hemodynamic-suite, MAP, shock index, cardiac-output math) but none of the dynamic indices. All three home in Clinical Math & Conversions (Group E):

Each tile reports the index, not the fluid order (spec-v11 §5.3): the give-fluid / withhold / start-pressor decision stays with the clinician and local protocol. The applicability caveats are rendered as a spec-v50 §3 posture note, not enforced as a refusal — PPV/SVV require a regular rhythm, controlled ventilation, and an adequate tidal volume; the IVC index mode must match the breathing mode; the PLR technique (semi-recumbent baseline, measure within ~1 min) is stated. Every ratio guards its denominator strictly positive, and a min above max (IVC/PPV) or a peak below baseline (PLR) is surfaced as a correctly-signed negative value in words, never clamped or leaked as NaN/Infinity. All three are Class A (fixed ratio arithmetic with cited thresholds), so none forces a docs/citation-staleness.md row. lib/fluidresp-v113.js + views/group-v38.js. Wave 3 reaches 495 (492 → 495, +3).

spec-v112 — ICU mortality & sepsis-coagulopathy (+5 → 492, Wave 3 opens)

Wave 3 (Critical care & pulmonary) opens with five standard bedside instruments. The catalog carried the ICU admission-severity models (apache2, saps-ii) and the sepsis-triage tools (sirs, qsofa-sofa, curb-65), but the front-door sepsis-mortality, sepsis-coagulopathy, ventilator-pneumonia, resuscitation-endpoint, and ICU-weakness instruments were reachable nowhere. Four home in Clinical Scoring & Risk (Group G); lactate-clearance is a Group E clinical-math tile:

Three source-governance catches (the spec-v97 re-fetch rule, three research agents cross-verifying each point table against ≥ 2 independent sources): (1) sic-score enforces the platelet + PT-INR subscore floor of ≥ 3 (not the ≥ 2 a first reading assumed) — so the SOFA item alone can never diagnose SIC; (2) cpis-vap uses the Pugin band-forms ≥ 50% leukocyte bonus, not MDCalc's absolute "≥ 500" rendering, and the ARDS exclusion on the oxygenation item is load-bearing; (3) mrc-sum-score grades elbow flexion (two secondary sources transcribe "extension" in error) and the ICU-acquired-weakness threshold is strictly < 48. lactate-clearance guards division by zero (initial must be > 0) and reports a rising lactate as a correctly-signed negative clearance, never NaN/Infinity. All five are Class A (fixed point weights / thresholds / arithmetic), so none forces a docs/citation-staleness.md row. Each renders the spec-v50 §3 clinical-posture note: it frames risk or likelihood, it does not author a resuscitation, anticoagulation, ventilator, sedation, or weaning order. lib/critcare-v112.js + views/group-v37.js. This opens Wave 3 of the spec-v100 program (487 → 492, +5).

Billing & reimbursement: what Medicare pays, whether the line survives, how the visit codes, what the drug bills, what the patient owes, and whether the claim is clean (spec-v77 → spec-v83, program complete)

The catalog has always been strong on the clinician at the bedside and competent on the operations clock (appeal/timely-filing/PA deadlines, the 2021 E/M selectors). What it lacked is the math a revenue-cycle professional redoes for every claim and cannot Google: what a line will actually pay, after every reduction. spec-v77 charters a six-spec billing & coding program and a new home group — Group B "Billing & Reimbursement" — governed by the same determinism bar as a creatinine-clearance calculator. spec-v78 ships the first feature: the MPFS reimbursement engine, five calculators (lib/billing-v78.js, views/group-b.js). Catalog 337 -> 342.

The Medicare allowable for one professional line is not a lookup — it is a computation, then a chain of reductions applied in a fixed order. The engine encodes the order once and each tile states where it sits:

   rvu-payment            mppr               bilateral-pay      multi-surgeon-pay   sequestration-adjust
   ───────────            ────               ─────────────      ─────────────────   ────────────────────
   [wRVU·wGPCI            100% of the        modifier 50 by     16% assistant /     2% of the program-
  + peRVU·peGPCI    ─▶    highest line,  ─▶  BILAT SURG     ─▶  62.5% co-surgeon ─▶ payment portion
  + mpRVU·mpGPCI]         50% each           indicator          / team by report   (after cost-share),
     × CF                 subsequent         (150/100/200%)                          never the cost-share
   = allowed (NF & F)     (endoscopy:        or a hard          gated 0/9 =          = net Medicare check
   + site differential     base rule)        not-payable gate   not payable

Worked anchor (hand-checked to the cent): CPT 99214, National Average GPCI (1/1/1), CY2026 conversion factor $32.7442 → non-facility (1.92 + 1.5 + 0.13) × 32.7442 = $116.24, facility (1.92 + 0.69 + 0.13) × 32.7442 = $89.72, site-of-service differential (1.5 − 0.69) × 1 × 32.7442 = $26.52. Money is integer cents end-to-end, formatted once at the render edge through fmt() — no float toFixed leak, zero NaN/Infinity by construction (the spec-v59 safety contract). Indicators gate, they do not guess: every indicator-0/9 path is a hard "not payable / does not apply" message, never a silent $0.