Skip to main content
Glama
README.md
# Concordance 2.0

A deterministic verification engine. It checks what is true and hands you a **receipt you can
re-verify** — a verdict, the worked reasoning, and a permanent content-addressed seal. It
eliminates what is not the answer so that what survives stands on its own.

> Rebuilt like a watch (the watchmaker's discipline): one bounded, counted movement; serviceable by
> any competent hand; machine-honest — the accuracy *is* the finish. See
> [`docs/V2_DEFINITION.md`](docs/V2_DEFINITION.md).
> New here? Start with [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md) — the newcomer's map (every count dated and derived from the code).

## The floor — seven primitives

Every answer is a `(verdict, trail, seal)` — three things, each independently checkable.

1. **Claim** — a structured assertion `{domain, spec}` (or text the router resolves)
2. **Verifier** — a pure deterministic function `spec → verdict + worked trail` (no I/O, no LLM)
3. **Verdict** — `PASS | REJECT | QUARANTINE` + the worked math (the elimination trail)
4. **Seal** — content-addressed (SHA-256) record of `{claim, verdict, trail}`
5. **Ledger** — append-only hash chain of seals
6. **Ranker** — IDF/full-text retrieval over the keeping
7. **Surface** — the face over the shared foundation (`secular` reach / `witness`)

The measure: **reproducibility, zero false-positives on the published benchmark (docs/BENCHMARK.md), re-checkability.**

## One foundation, two surfaces

The foundation is the Bible ([`FOUNDATION.md`](FOUNDATION.md)) — Christ at the center. From it the
engine reaches the world two ways: a secular `.com` (the reach) and an explicit-witness `.org`.
One engine, one foundation, selected by config — never a fork.

```python
from concordance import EngineConfig
EngineConfig("secular").identity   # the reach — no religious wording
EngineConfig("witness").identity   # the witness — the foundation named
```

## Run it

```bash
PYTHONPATH=src python -m concordance serve          # the .com reach (secular) — API + site, same-origin
PYTHONPATH=src python -m concordance serve --surface witness   # the .org witness face
```
Then open `http://127.0.0.1:8000/` — bring a claim, get a receipt. (Data — the keeping, the WEB
Bible, Strong's — is generated by `tools/migrate_*.py` and gitignored; the engine runs without it,
just with less to retrieve.)

## The surfaces — reached by asking, not by menus

- **The Console** (`/`) — speak or type anything; the keeping answers, shown as a calm trail
  (door → verifier → seal → source), read as a page or heard aloud in the operator's own cloned voice.
  Crisis-first, store-nothing.
- **The parable** (`/parable.html`, and the WALK door) — bring a situation; receive the one of Christ's
  parables whose *shape* fits it — the story verbatim, with his own explanation where he gave one, and
  one next step. Found and cited, never generated; honest silence where no single parable fits.
- **The Reading Room** (`/books.html`) — a free library: read our books in full, or open your own
  `.epub`, which is read entirely in your browser and never uploaded (store-nothing).

## For agents — the MCP (the church for agents)

The engine is also an MCP server (`python -m concordance mcp`). Five doors, every answer
found/verified/cited/sealed, never generated:

- **CHECK** — `verify` · `audit` · `seal_fetch` · `find_verifier`: a claim in; the verdict, the worked trail, the seal.
- **FIND** — `search` · `card_get` · `lookup` · `define`: one map, one lookup, the source itself.
- **WALK** — `ask` · `discern` · `parable` · `coach_next`: a situation in, one next step out — the path, the fitting parable, the figures' words.
- **KEEP** — `identity_create` · `study_create` · `decks`: a sovereign key born on your device; your deck, your receipts.
- **WORD** — `read_passage` · `word_study` · `cross_references`: Scripture at depth.

## Status

The 2.0 engine is built end to end and live (narrowhighway.com/.org/.tv) — floor 7/7, the derivation
moat (0 false-positives), the tamper-evident ledger + precedent overlay, **93 verifier modules** across
**77 routable domains**, the witness overlay (theology / witness / scripture / canon), the shared corpus
+ ranker (**880k+ cards**), scripture ref-resolution + original-language word study, the axes grid, the
Console + the parable register + the Reading Room, an MCP server (**100+ tools**), a sovereign HTTP API,
and the site. **400 test files**; the deploy gate (benchmarks, live assay, front door, recall) runs on
every push with **zero false positives** and reverts on any regression. The domain router is model-free
and measured (orthogonalized basis; routing cross-talk and invariance tracked by `tools/dispatch_probe.py`
and `tools/hodge_probe.py`). A gated **collective-intelligence layer** (stigmergy, quorum, flock,
eldership, physiology, inhibition) and an **alignment gate** (the keeping tiered by agreement — occult
explicit-only, other religions/philosophy referenceable) run over it; see [`docs/WORLD.md`](docs/WORLD.md).
Sovereign, stdlib-first; heavy deps (`sympy`/`scipy`/`numpy`) are optional, lazy-loaded.

1.0 is frozen (`v1.0-frozen` in the Lighthouse repo) — the shape that taught us the floor.

TDQS

C2.6/5.0

Scored across 102 tools

Disambiguation3/5

Most tools target a distinct resource, and the unusually verbose descriptions do a lot of disambiguation work. But several genuinely confusable clusters remain: verify vs. audit vs. discern vs. kernel_gate, seal_fetch vs. candidate_get, attest_record vs. witness_profession vs. witnesses, and search vs. locate vs. study_find vs. cards_browse. An agent will misselect within these families without re-reading the long descriptions.

Naming Consistency3/5

There is a recognizable family-prefix convention (coach_*, mesh_*, group_*, shelf_*, playbook_*, candidate_*, steward_*) mostly in snake_case. However, ordering flips between noun_verb (cards_browse, badges_issue, group_create) and verb_noun (read_passage, verify, curate), and a large set of bare single-word names (kernel, harmony, canon, seeds, decks, witnesses, profession) breaks any predictable pattern.

Tool Count1/5

102 tools is an extreme mismatch for any single server surface, spanning verification, Bible reference, community/mesh, sovereign identity, coaching, and budgeting. Many tools are thin single-purpose steps (signable/submit pairs) that inflate the count well beyond what an agent can hold in context.

Completeness3/5

The Bible-study surface is remarkably thorough (read, word study, cross-refs, harmony, atlas, canon, narratives, commentary, index), and verification/community flows are covered. But the surface is read-heavy and append-only: there is no delete/update for studies or cards, no group_leave, and no removal path for most created objects, leaving notable lifecycle gaps across the sprawling domains.

Maintenance

ActivityActive
ResponsivenessResponsive