Skip to main content
Glama

🛡️ HORIZON SHIELD

Verifiable construction estimate auditing for AI agents

Don't trust the estimate. Verify it.

An MCP server that lets AI agents check whether a Japanese construction or renovation estimate is fair, against open data, and returns a result anyone can verify against Bitcoin (OpenTimestamps). No account, no key.

MCP Registry Transport: streamable-http Open data: JCCDB 526,128 · CC BY 4.0 Anchored: Bitcoin / OpenTimestamps Auth: none GitHub stars HORIZON SHIELD KIRA on Glama Smithery A2A Python SDK: our signature vectors merged

Verified by Probity Verified by Probity (github.com/probityai/probity): the A2A conduct walk's RFC 8785 canonicalization passes the agent-evidence-vectors corpus, latest passing run

In one paragraph. JCCDB (Japan Construction Cost Database) is an open dataset of Japanese construction and renovation costs, created by Toshikatsu Oga (大賀俊勝), who has worked on construction sites for 30 years, and published by The HORIZONs Co., Ltd. under CC BY 4.0. Version 5.1 (2026-10-04, DOI 10.5281/zenodo.23133068; all versions 10.5281/zenodo.22127751) holds 526,128 records: 95,403 line items and 430,725 observations from 88 Japanese public sources, each observation with its evidence URL. The United States counterpart is USCCDB, the United States Construction Cost Database (DOI 10.5281/zenodo.22979157, 2,849,829 observations). HORIZON SHIELD (https://shield.the-horizons-innovation.com) is the buyer-side service built on JCCDB that checks whether a Japanese renovation estimate is fair.

Peer-reviewed (October 2026). Toshikatsu Oga, "A Demand-Side Benchmark for Consumer-Facing Construction Cost Questions: Price-Figure Span, Output Consistency, and the Case for a Verifiable Reference Layer", Intelligent Infrastructure and Construction 2026, 2(4), 15 (MDPI, open access, CC BY 4.0). DOI 10.3390/iic2040015. Forty Japanese renovation price questions put to frontier LLMs: the going-rate range is a median 2.5x to 3x wide, every price figure inside one answer spans a median 10x, and no answer offered an over-charge threshold. Questions, raw outputs, harnesses and scoring code are public.

Fair-price answers for common buyer questions (Japanese)

Each row is one question a homeowner in Japan asks, the page that answers it with the range from souba-db 2.2.0, and a JSON evidence object with the same numbers, a comparison across variants, steps to check a quote yourself, and the dataset hash anchored in JIDEC entry 42. Index: evidence/index.json.

Question

Answer page

Evidence object

外壁塗装 30坪 相場 いくら

/souba/gaiheki/

p002.json

外壁塗装の見積もりで150万円は高いですか

/souba/gaiheki-150man/

j001.json

屋根 葺き替え 30坪 費用

/souba/yane-fukikae-slate-hiyou/

p005.json

この屋根修理の見積もりが適正かどうか知りたい

/souba/yane-check/

j003.json

給湯器 交換 費用 相場

/souba/kyutoki/

p013.json

給湯器交換で20万円は高いですか

/souba/kyutoki-20man/

j002.json

トイレ リフォーム 費用 目安

/souba/toilet/

p016.json

シロアリ駆除 費用 適正価格

/souba/shiroari/

p071.json


Related MCP server: japan-real-estate-intel

Independent evidence, as of 2026-10-08

What people who do not work for this project have measured, signed or reproduced. Every row links to something you can fetch and recompute. The last two rows are the counts that are still small, stated as plainly as the rest.

Every night the TSUNAGI board runs these outside implementations again, cloned from their authors' own repositories at their latest commit, against the corpora in this repository, and keeps every disagreement on the board. A reproduction is checked again each day, not only on the day it was made. Add your implementation with one pull request to tools/tsunagi/implementations.json.

What

Who

Check it

Merged the five s0-control vectors of our a2a-card-sign-v01 corpus into the official A2A Python SDK's own test suite. On every change, its CI now checks that one Agent Card signed by a2a-sdk, @a2a-js/sdk, a2a-go and a reference signer verifies, canonicalizes to the shared bytes, and that the same card edited after signing is rejected. The review asked us to drop the vector fields the tests do not read and our frozen production card; we did, and it was approved and merged

a2aproject/a2a-python, reviewed and merged by Iwaniukooo11 (2026-10-07)

#1286, the vectors on main, the tests

Wrote a NENRIN provenance verifier from our specifications without reading any of our source, reproduced the 5 verdict signatures and 21 digests of interop-v0, and reported the 5 places where the text forced a guess; we fixed the text (interop-v0.1). A second implementation, written from VERIFIER.md alone, matches 18 of 18, and the runner of kuangmi-bit cut the expectations for the 36 edge vectors now in the tree (interop-v0.2/edge)

kuangmi-bit, A2A Discussion #1631 and PR #30 (2026-10-04 and 2026-10-05)

INTEROP.md, #1631, PR #30

Reviewed the NENRIN crosswalk to the agent governance vocabulary against commit a2c1c0b4: checked the cited lines of worker.js, agreement_verify.py, key_succession.py and RESUME_SPEC_v1.md, and recomputed the run0002 MUSUBI settlement byte for byte offline (anchor b13a3869..., settlement 11c27fcf..., exit 0). Two review rounds moved three rows down (governance_attestation left out until a reader can verify a signed witness record, bilateral_receipt and baseline_revision to false_analog), and the file was merged as reviewed

aeoess, maintainer of agent-governance-vocabulary (2026-10-05)

the review, the merged file

Found that an unsigned {"action": ...} passed as approval of a conditional action in settle v0 to v1.3, and that no settle version let anyone but the principal approve. babyblueviper1 wrote the approver fields and 9 vectors with a reference verifier of its own; settle v1.10 adopts them, and both implementations agree 9/9

babyblueviper1, Issue #29 (2026-10-05)

the report, the vectors, settle_v1_10.py

Installed nenrin-verify 0.4.5 from PyPI on a fresh runner of pipavlo82's own: settle v1.10 passes all 14 self-test groups, the approver vectors agree 9/9, all three mutants are caught, all 22 MUSUBI modules pass, run0002 reproduces, and all 51 vendored files match the source commit byte for byte. In a clean Windows venv, pipavlo82 found that the Node twin of contract_v0 never ran; fixed in 0.4.6. pipavlo82 confirmed the fix on Windows, then completed the full 0.4.7 suite there (Windows 11, Python 3.12.10, Node 22.15.0): 22/22, exit 0, in 1596 s, run0002 unchanged, all 53 manifest pairs verified against 827afef8. In pipavlo82's own words, a reproduction of published synthetic tests and fixtures, not a full independent security audit or a live two-party contract

pipavlo82, Issue #29 (2026-10-05)

the report, the run, the fix, the Windows fix confirmed, the full Windows run

Walked our gate from its own host, served the record content-addressed on its own domain, and filed the same bytes to our ledger signed with the key that domain serves

babyblueviper1's agent, api.babyblueviper.com (filed 2026-09-28)

the copy, the filing

Performed and signed an execution under a contract both sides signed. It was anchored in Bitcoin block 969090 and settled within_grant, final, with no deviations; the settlement recomputes byte for byte on two machines

api.babyblueviper.com as contractor, execution 19c44a79… (settled 2026-09-30)

run0002, the result

Signed, executed and settled a MUSUBI contract in which neither party is this project. agent-passport.org (principal) and kuangmi-bit.github.io (contractor) met in Issue #31, signed the terms, anchored the execution in Bitcoin block 970480 and published the settlement: within_grant, final. Our settle v1.10 recomputes the parties' settlement.json byte for byte and settle v1.16 gives the same verdict; the parties' files are now the real vectors of the settle core specification

agent-passport.org and kuangmi-bit.github.io, contract d7118f28…

the parties' repository, the settlement, #31, the specification and vectors, IETF Internet-Draft

Wrote a second implementation of the MUSUBI settle core from SETTLE_CORE_v0.md and NOTES.md alone, with an AI coding agent under aeoess's direction, from an input copy without the expected outputs, and froze code and outputs before comparing: all 76 vectors agree on all eight core fields. A question from aeoess about approval signatures found two sentences that allowed a reading the reference does not follow; the text was clarified (revision 0.1) and three vectors added. The 25 implementation notes from aeoess led to revision 0.2: eight places where the text was silent or wrong, parent narrowing stated in full, 38 more vectors. Updated to revision 0.2, the aeoess checker agrees with all 117 vectors with no field left unassessed (4ca508f), and 10 further notes from aeoess led to revision 0.3 and 6 more vectors. Updated to revision 0.3, it agrees with all 123 vectors with no field left unassessed (c0f8f4f); we reran it at c0f8f4f: 123 of 123, every output byte for byte. Updated to revision 0.4, it agrees with all 128 vectors with no field left unassessed, the 123 earlier outputs byte identical (53cfafb, run by its author and not rerun by us; not blind to the new cases; it holds for revision 0.5, which changed no rule, reason or vector). Run by its author, who maintains a party to the contract behind real_p31_baseline; we reran it at 7f9a224 and got 76 of 76

aeoess (Agent Passport System), Issue #49 (2026-10-17)

the checker, the report, our rerun, revision 0.4 at 53cfafb, NOTES.md N13

Wrote a third implementation of the MUSUBI settle core from SETTLE_CORE_v0.md (revision 0.5) and the published vectors only, without reading the reference or the second implementation: all 128 vectors agree on all eight core fields, with Ed25519 through cryptography or through its own RFC 8032 verifier; kuangmi-bit also ran the reference on the same machine, 128 of 128. We reran it at bbe455c on both signature paths: 128 of 128. The two notes on where the text and the vectors pulled apart (6.0, a strictly signed record whose body is not strict; 4.3, parent narrowing) became revision 0.6, two sentences, no vector changed

kuangmi-bit, Issue #49 (2026-10-19)

the implementation, the report, NOTES.md N20

Walked the gate's A2A face twice from a Windows PC of pipavlo82's own, and found that /a2a answered with an endpoint claim naming a different URL (EXT-001, fixed the same day with served_by). The fix was then verified offline: the originally captured response fails the new client, and a copy with only served_by changed fails too. Those two walks are unsigned and anchored in entry 58, and pipavlo82 checked their proof against Bitcoin block 968914. On 2026-09-29 came another walk, signed with a key served at pipavlo82.github.io (7/7), so the observations are now bound to a key pipavlo82 controls. On 2026-10-07 pipavlo82 filed another signed walk with the same key (7/7): a reader can check its signature from the ledger alone, because GET /witness/<sha> returns the canonical bytes, the signature and the key, and that key is the one pipavlo82.github.io serves (anchored in ledger entry 73, Bitcoin block 970423, 2026-10-08). In pipavlo82's own words, one observation from a Windows machine with our reference client, with Codex assistance and the Crystal Receipt context disclosed; no pool admission and no independence claim. A signed walk is attribution; the re-verification pool quorum still needs a witness that answers at its own /a2a

pipavlo82 (2026-09-27, 2026-09-29 and 2026-10-07)

#27, FINDINGS_EXTERNAL.md, the key, curl -s https://ledger.horizonshield.dev/witness/07838de9bbcd913bf56a4b1a59ddeb27a7c45225a23a1af047b939a854f521b6, the 2026-10-07 record, curl -s https://ledger.horizonshield.dev/witness/747ecd97635e104e75185212cfb70e3cc37310ea1f41966a818bfb5b5d57df15

Reproduced, in an independent Go implementation, the @a2a-js/sdk canonical form of our pinned gate card byte for byte (6410 bytes), offline from our committed fixture. kuangmi-bit corrected the fix our upstream report proposed: empty values have to be removed recursively

kuangmi-bit, in a2aproject/a2a-go#445 (2026-09-28)

the offline run

Reproduced our signature vectors in a reader of astrogilda's own: the 13 unknown-field cases (s3) and the 5 dual-name cases (s4), under both readings, and matched the a2a-go column of our MANIFEST

astrogilda, Probity's evidence vectors (2026-10-02)

probityai/agent-evidence-vectors#34, #42

Verified the a2a-python fix candidate and our 11 donated tests without using our test files, with an ES256 verifier written by hand; the length and sha of all 13 vectors checked match the canonical_utf8_hex in a2a-card-sign-v01

kuangmi-bit, a2a-python#1287 (2026-10-01)

the run

Published a candidate verifier that canonicalizes the received JSON; its self-reported results match our corpus on every vector (s0 5/5, s1 8/8, s2 11/11, s3 13/13), now a pinned column of the MANIFEST

aeoess, a2a-python#1287 (2026-10-02)

the candidate, #1287

A two party agreement signed by both sides, with consent to publish inside the signed bytes; both signing keys match the keys each domain serves

this project and api.babyblueviper.com, record 5d3e62f1… (accepted 2026-09-28)

curl -s https://agreement.horizonshield.dev/agreement/5d3e62f1f0b139992dd15db07b4abc6288a8a824effc31da8300ebe1d6ae6e6c/report

Re-verified the gate's agent card signature with their own verifier: 10/10, and the digest they recorded, 7c3bcb5f9633…, is the one our signing tool printed. They found and fixed two bugs in their verifier to get there

Agenstry, an independent agent directory in Amsterdam (2026-09-28)

listing, the card

Scored the JIDEC ledger export under an outside ledger conformance spec and asserted it in the VLC-1 CI: L1 against the head stamped in our daily Bitcoin batch, L0 on the log alone since the VLC-1 1.4.1-draft corrigendum found our end marker was not itself chained (EXT-022). Fixed on 2026-09-29 (833f8048): the marker now carries its own entry_sha256, linked to the last entry by the same recipe, so a rewritten marker breaks a hash; the VLC-1 rescoring is not yet published. An issue we reported is recorded there as EXT-020

the VLC-1 specification's maintainers

VLC-1, THIRD-PARTY.md

Outside operators who used the gate to measure servers of the operators' own

5 real hosts in the 30 days to 2026-10-02 (the counter lists 6; one is a test name that does not resolve), the last on 2026-09-23 (UTC)

curl -s https://gate.horizonshield.dev/usage

Rows on the public register that are not ours

2 of 10: one still pending, one added anonymously with POST /watch and not measured yet

curl -s https://gate.horizonshield.dev/register

Counted every week by tools/adoption/count_adoption.py, last on 2026-11-16. Every number comes from a source you can read; one that could not be read says so instead of counting zero. The whole count: ops/adoption/latest.json.

What

Count

From

Independent implementations that reproduced our bytes or verdicts

19 rows by 7 authors (a2a-card-canonical-form 1, a2a-card-sign-v01 7, a2a-card-walk 1, agent-card-signature 1, musubi-approval-v2 1, nenrin-action-presence 2, nenrin-provenance 3, nenrin-witness-reader 1, trace-intake 1, vate-nenrin-s3-crosswalk 1)

tools/adoption/registry.json, each row with its public link

Reproductions of our code and integrations of our vectors or records, counted apart from the row above

reproduction 2 rows by 2 authors, integration 1 row by 1 author

tools/adoption/registry.json, rows with kind reproduction or integration

Outside domains that signed a walk and filed it to the ledger

7 (agent-passport.org, api.babyblueviper.com, invinoveritas.dev, kuangmi-bit.github.io, ovaic.org, pipavlo82.github.io, sencor.ai)

every nenrin-witness-batch-v1 entry on the ledger

Re-verification pool

2 member(s) in 2 control cluster(s), 2 clusters needed for a quorum

workers/hs-ledger/nenrin/recovery-v0/pool_report.json, clusters from workers/hs-ledger/nenrin/recovery-v0/witness_diversity_report.json

MUSUBI contracts signed with an outside party

3 (with no party from this project: 1)

the signed contracts in workers/hs-ledger/nenrin/musubi-v0/, and contracts the parties publish themselves, listed in registry.json and signature-checked

...of which settled final, the parties' settlement recomputed here byte for byte

1

the settle verifier of the version each settlement names, run on the parties' own contract, event and header view

Outside identities that signed evidence (walk, contract or agreement)

7

the three rows above and the agreement records

Outside TRACE signing keys whose records were pinned with trace-pin-v0

0

every nenrin-trace-pin-batch-v0 entry on the ledger and the pending pool

Public repositories created from conduct-witness-template whose reproduce run succeeded in the last 30 days

1

GitHub API

Open: A NENRIN provenance verifier by a further author, in any language, that reproduces the verdict signatures of interop-v0, interop-v0.1 and interop-v0.2/edge. Two so far, both by kuangmi-bit: workers/hs-ledger/nenrin/interop-v0/INTEROP.md; Point any NENRIN provenance verifier, in any language, at the nine survivors vectors and post the result: npx nenrin-tsunagi run nenrin-interop-v0.3-survivors -- " {in} {out}". The reference takes one command after npm install nenrin-verify@0.5.2 (papers/nenrin-stack/README.md, section Rerun it from outside; replication package doi:10.5281/zenodo.23344510)

One external witness is a start, not a network. The re-verification pool below counts it as one control cluster, and a quorum of two independent controls is not met yet. Issue #27 is the open call for the second.

When every signature verifies: what can still deceive this system

Breaking a parser, a signature or a chain is the attack this repository was built against first. The harder question is the one left when all of that holds: every byte verifies, and the system is still told something false. These are the six ways we know of, what is built against each, how much of it real use has exercised so far, and what no amount of code here can do.

Attack

What is built

Exercised in real use so far

What it cannot do

Organizational Sybil: witnesses or parties that hold different keys but answer to one controller

witness_diversity.mjs groups the pool into control clusters, using the registered domain, IP, name servers, key and declared legal entity as strong signals and the ASN or shared hosting as weak ones, and the draw never takes a second member of a cluster. independence_v0.py counts the distinct legal entities among a contract's actors, from declarations signed by each key, against a quorum the contract states

The re-verification pool has one member, so it reports single_control_cluster and quorum short. Neither signed contract has stated an independence quorum yet. Since 2026-10-02 a contract can make finality depend on one: settle v1.8 refuses final until the stated quorum is met, and convergence v0 (settle v1.9) does not count a measurer that shares a signed history with a party. No signed contract uses either yet

Stop a Sybil that uses different providers and different legal entities. It makes one expensive and visible; it does not make one impossible

Signed lie: a correct signature over an observation that is false

corroboration_v0.py counts measurements by legal entity, not by signature, and counts one only when it sits inside a Bitcoin beacon and anchor window. Each item comes out corroborated, disputed, contradicted, undetermined or not corroborated. NENRIN keeps disagreements instead of resolving them, and response-v0 lets the measured party answer on the ledger

Two outside witnesses have walked the gate (issues #25 and #27), and one outside finding changed the protocol (EXT-001). No signed contract has required corroboration yet. Since 2026-10-02 one can: settle v1.8 holds final until corroboration is met, and settle v1.9 counts only measurers drawn by a future Bitcoin block from a pool both parties pinned, measuring after the draw, agreeing from two or more methods

Decide who is right when independent entities disagree, or catch a lie that every independent measurer tells

Semantic contract: both sides sign the same bytes and read "done" differently

terms_v0.py: every deliverable is an item of a content addressed vocabulary, with a quantity, a unit, a tolerance and a completion test. Free text is refused as unpinned_term, and check_completion is a pure function of the terms and the measurements

Both signed contracts are witness walks: done means a filed walk, which settle checks as within_grant. No contract has used vocabulary items yet. Under settle v1.8 a contract that opts into the gate must pin its terms, or it is not settled at all

Fix any meaning the vocabulary does not name

Hidden decision influence: a published rule, with a prompt, reward or incentive nobody sees steering the choice

The sieve and the policy behind a contract decision are signed and pinned by sha256 beside the contract (run0001). conduct-v1 requires an agent to disclose who pays it: referral, listing and success fees, in its agent card

In run0001 the sieve and the hidden instruction detector are pinned but not published, so the outside recompute marked them NOT CHECKED

Prove that nothing undeclared influenced a decision. A pinned rule shows which rule was applied, not that it was the only influence

Durable availability: the anchor survives and every copy of the bytes is gone

mirror-v0 lets anyone hold a copy and verify it offline against the anchors. Two honest mirrors share content_sha256. evidence-v0 copies the gate's verdict bytes into this repository every day and asks Software Heritage to archive it. /evidence/trace and /record serve raw bytes whose sha256 is their name

One mirror is held outside this company: babyblueviper1's, against which one of our agreements was diffed on 2026-09-28

Guarantee that any copy survives. Availability is the count of independent holders, and today that count is small

Fake observable surface: every public face reports healthy while the real state is not

The gate calls real tools where the operator consents, not only health pages, and the day and the tool it measures are derived from a Bitcoin beacon neither side chooses (nenrin_instant.js), so a face kept only for measurement day has no day to aim at. Outside witnesses walk from the witnesses' own networks, and a disagreement between vantages is kept. Runtime attestations can now be pinned beside those observations: POST /evidence/trace checks a TRACE Trust Record's signature and anchors its bytes (trace-pin-v0)

Two outside vantages so far. No TRACE record from an outside party has been pinned yet

See past a surface that is equally false to every observer at every moment. External observation measures what is served, and says so

Key compromise: a stolen key makes signatures that are mathematically genuine

key-history-v1 (spec, served at /.well-known/key-history.json): every key has a status (active, retired or revoked) and a compromised_from time, which is the earliest the key could have been exposed. A signature by a revoked key is attributed only when a clock the operator does not control shows the bytes existed before that time. The JWKS never serves a revoked key

Four keys are listed: card, agreement, witness and operator. None revoked

Say when a key was really stolen. compromised_from is the operator's statement, and so is anything built on it

Physical-world oracle: everyone signs that the work was done, and it was not

A measurement in corroboration_v0.py carries the Bitcoin block hash the measurer saw and later an anchor, so it cannot have been prepared before the contract or dated after it. It is counted per legal entity, never per signature

No contract has required a measurement yet; the first one that pays for physical work is the first test. Since 2026-10-02 such a contract can require the measurers to be drawn rather than chosen (settle v1.9), so the parties cannot bring their own

Turn the physical world into proof. When every entity lies together, the ledger keeps exactly who said so, when, and against which terms

Network: nobody else takes part

Nothing in code. Every door is open, needs no key and costs nothing

Counted in the table above, including what is still small. /usage has counted who asked, separately from what was measured, since 2026-09-28 (requesters). From gate 0.4.21 on, the gate's own measurements (its scheduled sweeps) are left out of /usage, so its counts are lower from that version

Create participants. Only use does that

One verifier threads the contract rows together: spine_verify.py reads a contract from what it was agreed to mean (terms_sha256), through who agreed and who did it, to who measured done, in entities, inside a block window. The first three rows share a limit worth naming once: the verifiers exist and are tested against their own attacks, but a verifier nobody's contract invokes proves only that it would work. The next contract that pays for real work is the one that has to state an independence quorum, name its deliverables from a vocabulary and require corroboration. Since 2026-10-02 the contract itself can make that binding: settle v1.8 (requirements.spine) refuses final until every stage is in place, and settle v1.9 (requirements.convergence) adds measurers nobody chose. What is still missing is a signed contract that uses them, and one between two parties neither of which is this project.

All of MUSUBI also installs without a clone: pip install 'nenrin-verify[ca2a]' (0.7.0) carries musubi-v0/ byte for byte, settle v1 to v1.16 and settle core v0 with its 128 vectors included. musubi-verify --settle-core, musubi-verify --pilot51-set2 and musubi-verify --run0002 recompute them; QUICKSTART.md takes five minutes.

One timeline: what each record answers

The layers above are not separate products. Each one answers one question about the same piece of work, in order, and each is a record you can fetch and recompute without trusting us. None of them answers the question of the row below it, and the table says what each one does not prove.

Order

Question

Record

Verify it with

What it does not prove

1

What did the parties agree?

MUSUBI contract (a2a-contract-v0), a grant both sides signed: allowed, prohibited and conditional actions, approvers, finality depth

contract_v0.py

That anyone will do it, or that it is a legal contract

2

What does the contractor say it did?

Execution record (a2a-execution-v0), signed by the contractor, naming the contract by contract_sha256

settle_v1_10.py

That the act happened in the world

3

What did the runtime attest it ran?

A TRACE Trust Record, pinned by the sha256 of its RFC 8785 form and anchored

trace-pin-v0

That the attestation's own claims are true. No TRACE record from an outside party has been pinned yet

4

What did someone outside observe?

Witness record (jidec-path-v1), signed by the witness with a key its own domain serves

nenrin-verify, or replay the walk with a2a_conduct_walk.py

That any response was correct, or that the agent behaves the same at other instants

5

Did the observers agree?

Discrepancy record, kept instead of resolved, and the measured party's signed reply (response-v0)

NENRIN_DISCREPANCY_0001.md

Which observer was right

6

When it broke, did it really come back?

TSUGI recovery chain: drift, proposal, authorization, execution, re-verification by witnesses drawn from a Bitcoin block

tsugi-verify (recovery-v0)

That the repair was the right one, only that it was authorized, done and re-checked

7

Could any of this have been changed later?

JIDEC: every record above lands in an append-only, hash-linked ledger whose entries are stamped to Bitcoin

Verify this project without trusting it

That a record is true, only that it existed by a block and has not changed since

Settle reads rows 1, 2, 4 and 7 together: it recomputes whether the anchored executions stayed inside the grant, and from v1.8 on a contract can refuse final until independent measurements agree (row 4) and the measurers were drawn rather than chosen. The row no layer can fill is the one the section above ends on: when every independent party tells the same lie, the timeline records exactly who said it, when, and against which terms. Runtime evidence (row 3) and outside observation (row 4) are complementary for that reason: one says what ran inside, the other what was served to everyone else, and a contract can require both.

Price ranges you can recompute

Every get_price_range answer carries a recompute block: the URL and SHA-256 of the souba-db.json bytes it used (the same file is in this repository at data/souba-db.json), the entry id of each row, and the formula: the table value, or the table value times the regional multiplier in the same file, rounded half up. tools/recompute_price_range.py checks an answer end to end with the Python standard library, and fails on any byte or row that differs.

What it shows is that the answer equals the published table. It does not show that the table is right: the values are curated by a named curator against the sources listed in the file, not computed from those sources by a published formula.

NENRIN: tree rings for AI facing services

A tree adds one ring a year. Nobody can paint one in afterwards. NENRIN gives that property to software services.

In the thirty days to 2026-08-14, this server's listing on one MCP directory was shown 93,983 times, by that directory's own dashboard (24 clicks, 421 profile views, and zero tool calls made through the listing). Those are impressions on one directory, not search results across search engines or AI assistants. Calls that reach the server by other routes are a different number, and how many of them came from outside, we cannot say: the usage counter deliberately stores no IP addresses, so it cannot separate our own automated checks from external traffic. An earlier version of this paragraph said the server was invoked 0 times and called the 93,983 AI search results. The zero holds only for calls through that one listing, so the claim for the server as a whole is withdrawn here rather than quietly deleted, and the NENRIN specification's copy of the sentence is corrected by a later ledger entry (seed_entry_nenrin_spec_correction_20261013.json) because the anchored bytes cannot be edited. Discovery is solved. Choice is not. An agent picking between 90,000 servers can only read what each vendor wrote about itself. NENRIN adds the missing layer: records of conduct that the vendor did not author and cannot delete.

How it works, in three lines:

  1. Open witnessing. Anyone can measure any endpoint and submit the walk to the public ledger under the witness's own name and vantage. The operator holds no veto: acceptance is mechanical schema checking, and the code that enforces this is in this repository.

  2. Discrepancies are the product. When two witnesses report incompatible observations of the same target, the disagreement itself becomes a permanent, citable record. The founding one is real: NENRIN_DISCREPANCY_0001, two honest witnesses, one target, both correct.

  3. Rings. Each month the accepted records bundle into a ring that carries the hash of the previous ring, timestamped to Bitcoin. Eighteen months of rings cannot be created in an afternoon, by anyone, including us.

The specification is anchored on the public ledger as entry 19 (sha256 9ccba2e325fd2a555fcdb2dec519b8c6bf7a669064674846aea98ecfff824e3d): NENRIN_SPEC_v1.md. It names its own prior art (Certificate Transparency, Rekor, in-toto, SLSA, OpenTimestamps), states exactly which combination is claimed as new, and invites refutation into the same ledger.

The witness intake is live. Start here:

curl -s https://ledger.horizonshield.dev/witness

We are the first test subject under our own rules. The ledger keeps the record of our gate failing its own test, and the full 522 incident that started all of this. Unflattering records stay.

If a register that cannot delete criticism of its own operator is infrastructure you want to exist, star this repository. Stars are how researchers and agent platforms find it. The rings accumulate either way. They accumulate faster with witnesses.

Task-bound conduct: binding an A2A Task to its evidence

Choice does not end when an agent picks a server. It picks, then it delegates a task. What that delegated task actually did, observed by someone other than the two parties, is the evidence the next agent needs. NENRIN binds it to the A2A Task id itself (a2a.task.id, aligned to A2A issues #1769 and #2103), not to "this server failed once".

The conduct walk already talks to a real agent over A2A and receives a real Task with an id. It now files a signed, content-addressed observation bound to that id: who delegated to whom on task T, and how the walked agent behaved. The witness signs the observation (witness_sig), the requesting party signs the delegation edge (edge_sig), and both keys live inside the did:key identifiers, so anyone verifies with no network and no trust in us.

Three reads, each a real record you can fetch now:

curl -s "https://ledger.horizonshield.dev/witness/task?task_id=d1651c71-28b0-422b-8f14-e2dc66c5a145"
curl -s "https://ledger.horizonshield.dev/trust-signal?task_id=d1651c71-28b0-422b-8f14-e2dc66c5a145"
curl -s "https://ledger.horizonshield.dev/witness/task/evidence/0bff13042d89ec0bc33f2f5149612774f8a706e67c1f345abfc1f10962e31608"

The first returns the full witness set per delegation hop, with the aggregate verdict computed so a disagreement is preserved and never the favorable one. The second returns the same as a consumable signal that carries counts and verdicts and never a numeric score. The third returns one observation with its anchor status: this evidence sits in NENRIN ledger entry 44, a nenrin-task-witness-batch-v1 bundle timestamped to Bitcoin like every other ring.

What the signatures prove, stated plainly: who asserted the observation and who attested the delegation edge, not that the assertion is true. The ledger attests that the witness is distinct from both hop parties by key (R1); it does not attest operator independence, so a self-witness satisfies R1 and says so in its own record. A genuine third-party observation is the same walk run by someone with no stake, filed to the same live endpoints.

The loop this closes: an agent discovers a server, reads conduct the server did not write, chooses, delegates a task, the task is witnessed, the evidence accumulates bound to the task id, and the next agent chooses on it. The code is in workers/hs-ledger/nenrin/task-delegation-bind-v0 (the ledger faces and the producer) and workers/hs-ledger/nenrin/a2a-conduct-walk (the walk that binds, with --bind-task).

TSUGI: proof of recovery, the second pillar

Verification says whether an endpoint conforms today. It says nothing about what happened when it broke, or whether it is really back. TSUGI (継, from kintsugi: the repair is visible and becomes part of the object's history) is the layer after verification. It does not repair; it proves recovery.

Five record types, hash-linked: drift (a witness measured a public surface and it did or did not match), proposal (one repair from a closed catalog of five primitives, with what it does not establish), authorization (the operator's Ed25519 signature over the proposal hash, expiring), execution (before and after state), verify (the witness measured again). The verifier refuses an execution of a human-approval primitive whose authorization is unsigned, signed by an untrusted key, or expired. The operator's public key is served at https://gate.horizonshield.dev/keys/operator.json, the same way the agreement and witness keys are.

Re-verification witnesses are not chosen by the operator. They are drawn from a public pool with sha256(bitcoin block hash | pool hash | record hash) as the seed, so a third party recomputes who should have been asked. A drawn witness receives only a blind request (no expected values) and returns only a signed observation; nothing it says is executed. This is the July 2026 lesson turned around: unknown agents may observe you, never instruct you.

Two real incidents are recorded in workers/hs-ledger/nenrin/recovery-v0: a raw deploy that bypassed the deploy guard and silently broke the card signature and the OpenAI domain challenge (found by an external verifier, closed with an unsigned chat approval, which the strict verifier flags as such), and a card signature broken by two version bumps deployed without a re-sign (found by the daily witness, closed with a signed authorization, twelve records). The pool of external witnesses was empty on 2026-09-20, and the records say so instead of pretending a quorum. It held one member, babyblueviper1's agent, until 2026-10-22, and the records reported the quorum as short rather than met. Since commit 89305502 it holds two, api.babyblueviper.com and kuangmi-bit's CloudBase host, which the diversity check (witness_diversity v2.3) counts as two control clusters, so a quorum of two can be met. Incident 2 can be recomputed in a browser, hashes and the operator's Ed25519 signature, with no trust in this project: https://shield.the-horizons-innovation.com/tsugi/ Its chain file and record hashes are anchored as JIDEC entry 50 (OpenTimestamps, Bitcoin).

Repository map

Path

What it is

workers/hs-verify-gate

The verification gate: nightly sweeps, on demand checks, probed_via route disclosure, gate_commit pinning, surface change tracking

workers/hs-ledger

The JIDEC append only ledger and the NENRIN witness intake

workers/hs-ledger/nenrin/task-delegation-bind-v0

Task-bound conduct: an A2A Task id bound to a signed, Bitcoin-anchored witness observation; the /witness/task, /trust-signal?task_id and /witness/task/evidence faces

workers/hs-ledger/nenrin/a2a-conduct-walk

The conduct walk that measures an agent and, with --bind-task, files the observation under the real a2a.task.id

workers/hs-ledger/nenrin/agreement-v0

The agreement record: two agents, two signatures, one set of bytes. Verifier written twice, in Python and JavaScript, and proved to agree

workers/hs-ledger/nenrin/recovery-v0

TSUGI (継), the second pillar: proof of recovery. A drift witness measures eight public surfaces of the gate daily; a repair is proposed from a closed catalog, authorized with the operator's Ed25519 key (trust anchor at /keys/operator.json), executed, re-verified, and the whole chain is hash-linked so anyone recomputes it. Random re-verification witnesses are drawn from a public pool with a Bitcoin block as the seed, so nobody can claim the operator chose them. Two real incidents are in the tree as 7- and 12-record chains. The verifier is JavaScript (recovery_verify.mjs, one file as sdk/tsugi_verify.mjs); the repository's Python twin (recovery_verify.py) agrees with it on the record bytes, the hashes, the signatures and the draw, and on the refusal codes its twin test checks. The full report, refusal text included, is matched by the Python package nenrin-verify (0.2.0, tsugi-verify): the same output as node tsugi_verify.mjs, byte for byte, on 98 frozen cases and on 29,228 edits of the real chains

workers/hs-verify-relay

The public edge relay born from the 522 incident (documented in the discrepancy record)

verify-directory

The public register page: every listed server, our own included, with its live verdict

everything else

The GitHub Pages site for the human facing service at the-horizons-innovation.com

The agreement record: the other half of a measurement

A conduct record is one sided. Somebody measured somebody. Nothing in it records the other half of commerce: that two agents agreed on terms, and that both said so.

a2a-agreement-v1.1 is that record. At time T, party A and party B both signed the same canonical bytes describing terms, and each of them pinned, by sha256, a conduct record about the OTHER party written by somebody who is neither of them.

What it refuses to be is as load bearing as what it is. No custody. No matching. No editorial step. The recorder must not hold funds, must not decide whether a deal happens, and must not charge a fee that varies with the amount or the outcome. A record whose fee moves with the number is refused by name. Refusal is mechanical, and none of the terms are ever judged by anyone in this layer.

The claim is not a new primitive. It is the combination: two mandatory signatures, the counterparty's measured conduct pinned by sha at the moment of signing, an intake that judges nothing, and an external anchor nobody here operates. Prior art is named in the draft rather than left for a reader to find: AP2, x402, ACP, MPP, Cedulon, the 1F916 Agent Record, and SCITT.

The verifier is written twice. Once in Python, once in JavaScript, by design and not by accident: two implementations that disagree are the exact seam this project measures everywhere else, and building one into this layer on purpose would be a poor joke. 5,286 frozen cases, and the two produce the same report byte for byte, including every refusal code and the English sentence attached to it. Proving that moved the Python once, when the JavaScript disagreed on two cases and the check that settled it was running the Python against its own frozen fixture, where it failed the same two.

Then the rules were broken on purpose, 77 ways in Python and 36 in JavaScript, to find out whether the 5,286 cases could tell. Six breakages survived, and not one was a defect in either implementation. They were holes in the test set. All six are closed.

The intake exists since 2026-09-16 (agreement_intake.mjs, wired into the ledger worker): POST /agreement accepts a record only when both signatures verify against the keys each party serves at its key_url, deduplicates atomically, serves the record by sha at GET /agreement/{canonical_sha256}, and bundles the accepted pool into a daily anchored ledger entry. It judges nothing. The first record it accepted, between this project's agent and babyblueviper1's agent, is in the tree as first_agreement_record.json. An earlier version of this paragraph said there was no intake; the code had been written and deployed but not committed, which this repository noticed on 2026-09-20 and corrected.

The second record between the same two agents was re-signed on 2026-09-28 with "publication": "public" inside the bytes both parties signed (record-privacy-v1: nothing bilateral is published on one side's say so). It was accepted the same day with both signing keys matching the keys each domain serves, and each side's pinned conduct record is filed on the ledger: 5d3e62f1…/report.

The register, as a repository

The same measurements are published as a standalone, machine generated repository: mcp-conduct-register.

Nobody selects the rows there either. A script rebuilds the table from the public API once a day, and the same run writes a register.json snapshot so an agent can read the register without parsing Markdown. It carries a CITATION.cff, so the register can be cited the way a dataset is cited, and an llms.txt that states in plain words what the register is and, more importantly, what it is not.

Three ways in, none of which need us

Since 2026-09-04 the gate can be used without asking anyone at HORIZON SHIELD.

For the server you operate. Put {"allow_tool_call": true} at /.well-known/mcp-conduct.json on your origin. Only the owner of an origin can place a file there, so the gate takes it as consent, measures determinism on the public register with it, and writes into every verdict where it read it (gate 0.2.4). Add a compensation block to your agent card (paid_by, referral_fee, listing_fee; the content is not judged, only its absence) and POST /watch once. A row can then reach verified with no hand of ours involved.

For your CI. One step measures the server on every push and recomputes the verdict hash on the runner, so the gate is never trusted: wedjat-check-action (uses: ogasurfproject-jpg/wedjat-check-action@v1). It fails the job on a measured failure and leaves unmeasured conditions unmeasured; require and must_pass decide how strict that is.

For the agent that connects. mcp-conduct on npm (zero dependencies) reads /is-verified before an MCP client connects and applies a policy you choose: warn, measured (block only what was measured and did not pass), verified-only, or off. verified is true or null, never false; not measured is never failed. Source: mcp-conduct.

Stated plainly: as of 2026-09-05 the register holds our own servers and nobody else's. The doors are open; the first outside row has not walked through yet.

JIDEC: verify this project without trusting it

The verification process behind HORIZON SHIELD's results is published as a Bitcoin anchored, append only public ledger. You do not have to trust us: fetch the anchored bytes, hash them yourself, and check the timestamp.

One line is enough to check any entry:

curl -s "https://ledger.horizonshield.dev/ledger/5?format=raw" | shasum -a 256

What this proves and what it does not is stated by the ledger itself at /health under transparency, including that OpenTimestamps has no RFC, ISO or eIDAS standing.

The previous hostnames, hs-ledger.oga-surf-project.workers.dev and hs-jidec-mcp.oga-surf-project.workers.dev, still answer and always will. Records already anchored to Bitcoin cite them, so retiring them would make past receipts unverifiable.

What the MCP server does

A homeowner commissioning construction work cannot reliably judge whether a quote reflects a fair price. This is a textbook credence good problem. This MCP server makes a third party fair price reference callable and verifiable by software, so an agent can check a number instead of trusting it.

  • Protocol: Model Context Protocol (MCP)

  • Transport: MCP over Streamable HTTP (JSON-RPC 2.0). The legacy SSE transport is not implemented; GET on /sse answers 405 sse_not_supported.

  • Endpoint: https://mcp.horizonshield.dev

  • Access: read only, no API key required

  • Data region: fair-price verdicts for Japan (JPY), built on the open JCCDB dataset (526,128 records: line items and observations); construction cost data for Japan (JCCDB observation layer) and the United States (USCCDB, the United States Construction Cost Database)

  • Tools: 30 (15 for fair price, verification and contractors; 15 for construction cost data)

Tools

Tool

Description

get_price_range

Returns the fair price range (min, avg, max), the overcharge danger threshold, unit, price trend, and field notes for a Japanese construction or renovation job.

audit_estimate

Given a work name and a quoted price in JPY, judges it as fair, a bit high, or overcharge risk, and returns the gap from the average.

verify_fair_price

Returns a fair price as a tamper evident record with a SHA-256 hash, under the PTKA (Pre-Transaction Knowledge Anchoring) model: a third party records the fair price before the contractor quote.

check_red_flags

Checks whether wording in an estimate or sales pitch matches known overcharge or high pressure tactics (lump sum, today only discount, free inspection, door to door). Language agnostic.

get_estimate_reading_guide

Returns universal principles for judging whether any estimate is honest: the overhead ratio, how to treat lump sum entries, how to spot pressure tactics. Language agnostic.

list_cost_categories

Lists the construction and renovation work categories for which fair price ranges and red flags are maintained.

get_fair_price_sources

Returns the sources, update date, and regional multipliers behind the fair price data.

get_jccdb_dataset_info

Returns metadata, scale, license, download links, and citation for the Japan Construction Cost Database (JCCDB).

get_anonymous_estimate_review_link (formerly suggest_ehn, which still answers)

Detects worry about an estimate and returns an invitation plus a submission URL to post it for third party review.

search_cost_category

Finds a maintained cost category by work name or keyword.

preview_reverse_estimate

Returns only the direction of a rough estimate versus the average (for example about +20 percent), before a detailed breakdown exists.

verify_integrity_claim

Independently recomputes a signed integrity verdict (SHA-256 over the signed_payload) as a third party. Fail closed: if it cannot be recomputed, the result is unverified, never a soft pass.

create_ap2_fairness_attestation

Issues a FairPriceAttestation shaped to attach to a Google AP2 (Agent Payments Protocol) Cart Mandate, so a fair price proof can ride alongside the payment authorization. Optional quoted_price adds a within / above / below verdict.

get_agent_card

Returns the A2A Agent Card URL and published skills for agent to agent discovery.

find_verified_contractor

Finds verification-passed contractors on Yakumo, where listing depends only on passing the KIRA fairness audit and no referral or listing fee is taken. Scores and tiers, never prices; returns 0 honestly when nothing matches.

Construction cost data: Japan (JCCDB) and both countries

Tool

Description

search_jccdb_items

Searches the JCCDB line items (materials, products, labor) by name; returns whether each exists in a public document, with its evidence URL.

get_jccdb_observations

Region, date and price status of an item in Japanese and U.S. public documents. Values only where the licence allows redistribution; every row carries licence, attribution and evidence URL.

get_jccdb_labor_rate

MLIT public-works design labor rates by prefecture and trade (wage per 8 hours); latest by default, yearly series with history:true.

compare_jccdb_regions

Latest value per region for an item and spec, with min, median (computed) and max; only identical spec, unit and basis are compared.

get_jccdb_work_unit_price

Public-works unit prices for work items (materials, labor and equipment combined), with composition-ratio rows. Not renovation quote prices.

get_jccdb_index_series

Construction cost index series (NHCCI, PPI, MLIT deflator and others) over a period, with year-over-year change computed by this service.

get_jccdb_coverage

What the observation layers hold: rows per country, layer and source, priced rows and source periods; empty combinations are listed as absent.

Construction cost data: United States (USCCDB)

USCCDB is the United States Construction Cost Database: U.S. public-domain federal data and city open data, one row per observation with source URL, sha256 and licence. The four chain tools compute on request and are not distributed as files. Public-works prices, statistics and estimates are reference data, not renovation quotes.

Tool

Description

get_us_construction_prices

U.S. public construction cost data by layer, region, period and item: Davis-Bacon wages, BLS wages, public unit costs, equipment rates, permits and spending, indexes and area factors, HUD cost limits, state DOT bid prices.

get_us_prevailing_wage

Davis-Bacon general wage determinations by state, county and trade: base wage and fringe with decision number and source URL. Minimums for federally funded work, not private market rates.

get_us_permits

U.S. building permits by region and year: Census BPS and distributions of declared valuations in city permit data. Not contract prices.

get_us_area_factor

DoD Area Cost Factors and USACE CWCCIS state adjustment factors by state, county, ZIP, city or overseas country. Budgeting factors, not a test of a quote.

get_us_price_chain

Estimated U.S. prices along the distribution chain for construction materials and chemicals: landed import cost, wholesale, retail range and contractor, with formula, source URL and sha256 on every row. Computed on request.

get_us_import_landed_cost

Landed cost of U.S. imports by HS 10-digit code (Census IMDB): customs value, CIF, calculated duty including Section 232, unit cost, effective duty rate and top partner countries.

get_us_trade_margins

U.S. wholesale and retail gross margins by NAICS (Census AWTS, ARTS, AIES 2024) with kake_cost_ratio = 1 - margin; optionally the BEA 2007 margin structure. Industry averages.

get_us_contract_discounts

Published discount rates off list price in U.S. public contracts (Washington DES, NASPO ValuePoint MRO) with kake_ratio = 1 - discount. Rates are ceilings; list bases differ by row.

The same public data, read for a homeowner holding a quote, one question per page with every figure linked to the record it comes from: Guides · Roof replacement cost in Austin, TX · Is my roofing quote too high? · Overhead and profit on a contractor quote · Trades hourly rates in Austin, TX · How many squares is my roof? · Do I need a roof permit in Austin? · Can a roofer waive my deductible in Texas? · Price good only today · How much deposit? · Material prices up: is the quote justified? · Lead paint and RRP certification · Scope of work before quotes. The quote check itself: shield.the-horizons-innovation.com/us/.

Connecting

This is a remote MCP server. Point any MCP client at the endpoint.

{
  "mcpServers": {
    "horizon-shield": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://mcp.horizonshield.dev/"]
    }
  }
}

If your client supports remote MCP servers directly, use the endpoint URL above.

In Claude, without configuration

The construction cost data server (https://ccdb.horizonshield.dev/mcp, fifteen read-only tools for JCCDB and USCCDB) is listed in the Claude connector directory after Anthropic's automated review: https://claude.ai/directory/connectors/horizon-shield-construction-cost-data . Open the page in Claude and press Connect; no key and no account with us. It is a community connector, which means it passed the automated review and is not verified by Anthropic.

Example

audit_estimate(work: "外壁塗装 30坪", quoted_price: 1500000)

Returns a verdict (for example, overcharge risk), the fair range (min, avg, max), and the gap from the average. verify_fair_price additionally returns a SHA-256 fingerprint of the fair price claim under PTKA (the PTKA declaration is anchored in Bitcoin; the receipt is kept by the issuer, and from 2026-10-19 each day's receipts are listed in one JIDEC ledger entry the next day).

Verify a verdict yourself

Every verify_fair_price call returns a verify_url of the form https://shield.the-horizons-innovation.com/verify/?id=<claim_sha256>. The public verify page recomputes the SHA-256 in your own browser (Web Crypto) and checks it against the receipt. Nothing is sent to any server. The same claim is served back as JSON at https://mcp.horizonshield.dev/ledger/<claim_sha256>. Trust is conferred by recomputation, not assumed in the issuer.

Twenty real overcharge diagnoses are also published as tamper evident receipts, each with claim.txt, its SHA-256 digest, and an OpenTimestamps proof:

sha256sum claim.txt
ots verify -f claim.txt proof.ots

AP2 bridge

Google's Agent Payments Protocol (AP2) makes what a user authorized verifiable through a signed, tamper evident Mandate. create_ap2_fairness_attestation issues a parallel attestation that makes value verifiable, shaped to attach to an AP2 Cart Mandate before the user signs. Parallel layers, same philosophy: pre transaction, tamper evident, independently recomputable.

Data and academic record

Author

Toshikatsu Oga (大賀俊勝), The HORIZONs Co., Ltd., Hiratsuka, Japan. A carpenter of thirty years. ORCID 0009-0000-9180-903X.

"Cheapest is not the same as fair."

"Verify, don't trust."

"Thirty years on site taught me the enemy is the middleman, not the craftsman."

Full collection (50 quotes, JSON-LD): TOshi Oga, in his own words

Live diagnostic: https://shield.the-horizons-innovation.com · The Evidence: https://shield.the-horizons-innovation.com/evidence-en/ · The Movement: https://shield.the-horizons-innovation.com/movement-us/

License

Data: JCCDB, CC BY 4.0. Server code: see the LICENSE file in this repository.

Available Tools

15 tools
audit_estimateAudit Estimate Against Fair PriceA
Read-only
Inspect

業者が提示した見積金額が適正かを、HORIZON SHIELDの適正レンジ(souba-db, 大賀俊勝 実務監修)と照合して判定する。手元に具体的な見積額がある時に使う。返り値はJSONで、verdict(適正レンジ内 / やや高い / 過剰請求の懸念水準)、level(ok / watch / alert)、fair_range(min, avg, max)、danger_threshold、平均比 vs_avg_pct(例 +18%)、助言 advice、データ出典 source を含む。工事名が見つからない場合、近い候補があれば did_you_mean として返す。単価(平米など)建ての工事に総額らしい金額を渡した場合は unit_mismatch の案内を返す。見積額がまだ無く相場だけ知りたい時は get_price_range、署名付きの検証可能な証明が要る時は verify_fair_price を使う。Japan only, JPY。 / Audits whether a contractor quoted price for a Japanese construction or renovation job is fair by comparing it against HORIZON SHIELD fair-price ranges (souba-db). Use when the user already has a specific quoted amount. Returns a JSON object with verdict, level (ok, watch, alert), fair_range (min, avg, max), danger_threshold, percentage gap versus the average (vs_avg_pct, e.g. +18%), advice, and data source. If the work name has no match, close candidates may be returned as did_you_mean. If the work is priced per unit and the amount looks like a total, a unit_mismatch notice is returned instead. For the typical range only use get_price_range; for a signed verifiable attestation use verify_fair_price. Trigger phrases: この見積もり高い?, 適正?, ぼったくり?, 妥当?, is this quote fair, am I being overcharged, is this a rip-off.

Inputs: work (required), quoted_price (required), region. Returns: 見積額の適正診断。verdict・level(ok/watch/alert)・fair_range・danger_threshold・平均比・助言・出典。 /…

check_red_flagsCheck Estimate Red FlagsA
Read-only
Inspect

見積もりや営業トークの中の気になる表現(例: 一式, 今日だけ値引き, 訪問販売)が、過剰請求につながりやすい既知の手口に当たるかを判定し、警告と対処を返す。代表的な手口のみを判定する。 / Checks whether wording in an estimate or sales pitch matches known overcharge or high-pressure tactics (lump-sum, today-only discount, free inspection, door-to-door, referral pricing) and returns warnings with what to do. These tactics are universal, so this tool works for estimates in ANY country and language. Covers representative tactics only.

Inputs: text (required). Returns: 既知の過剰請求・強引営業の手口との照合結果。該当した手口と警告・対処。 / Matched overcharge or high-pressure tacti…

create_ap2_fairness_attestationCreate AP2 Fairness AttestationAInspect

このツールは決済を開始・承認・実行しません。資産・通貨・暗号資産の移動も行いません。発行するのは適正価格の証跡だけです。呼び出すたびに公開台帳へ記録を1件追加するため読み取り専用ではありません。 / This tool does not initiate, authorize, or execute any payment, and does not move funds, currency or crypto assets. It only issues a price-fairness attestation. Each call appends one record to the public ledger, so it is not read-only. AP2(Agent Payments Protocol)対応エージェント向けのブリッジ。決済カート(Cart Mandate)に添付できる適正価格の証跡(FairPriceAttestation)を発行する。AP2のMandateは『ユーザーがこの支払いを承認した』ことを検証可能にし、この証跡は『その価格が適正である』ことを検証可能にする。認可の検証と価値の検証、二つは並列レイヤー。quoted_price を渡すと適正レンジ判定(within/above/below)も同梱する。証跡は SHA-256 と公開台帳と verify_url で誰でも再計算検証できる。 / Bridge for AP2 (Agent Payments Protocol) agents: issues a FairPriceAttestation that a shopping or payments agent can attach to a Cart Mandate before asking the user to sign. AP2 mandates make authorization verifiable; this attestation makes value verifiable. Parallel layers. Pass quoted_price for a fair-range verdict (within, above, below). Independently verifiable via SHA-256, a public ledger and a verify_url. Japan construction and renovation pricing, JPY.

Inputs: work (required), merchant, quoted_price. Returns: AP2 Cart Mandate 向けの適正価格証跡。attestation(FairPriceAttestation)・cart_mandate_examp…

find_verified_contractorFind Verified Contractor (Yakumo)A
Read-only
Inspect

地域と工事名で、Yakumo(検証を通った加盟店だけが並ぶ建設モール)の検証済み施工店を探す。掲載は KIRA 適正診断の通過だけで決まり(fail-closed)、紹介料・掲載料は受け取らない中立の名簿。金額は出さずスコアとティアで示す。検証手続き中の店は pending として別に返す。条件に合う検証済みの店が無い時は 0 件と正直に返す(名簿は小さい)。価格の照会(get_price_range / audit_estimate)の後に、施主が『どこに頼めばいい』『信用できる業者は』と聞いた時に使う。 / Finds verification-passed contractors on Yakumo, a directory where listing depends only on passing the KIRA fairness audit (fail-closed) and no referral or listing fee is taken. Returns scores and tiers, never prices; pending stores are returned separately; returns 0 honestly when nothing matches (the directory is small). Use after a price check when the user asks who to hire or which contractor can be trusted. Trigger phrases: 業者を探したい, どこに頼めば, 信用できる工務店, find a contractor in Japan, who should I hire.

Inputs: area, work. Returns: 地域・工事名に合う検証済み施工店(Yakumo)。stores(検証済み)・pending_stores(手続き中)・directory_size・neutr…

get_agent_cardGet A2A Agent CardA
Read-only
Inspect

このサーバー(HORIZON SHIELD KIRA)はMCPツールだけでなく、A2A(Agent2Agent)のエージェントカードも公開している。外部のA2A対応エージェントから発見・連携したい場合の、エージェントカードURLと公開スキルの一覧を返す。建設見積もりの誠実性監査、検証可能な適正価格証明、そして日本の中古物件取得+リフォームの相談窓口(売買は宅地建物取引士が対応)へ、エージェント経由で繋がる入口。 / Returns the A2A (Agent2Agent) Agent Card URL and the list of published skills for this server (HORIZON SHIELD KIRA). Use when an external A2A-capable agent wants to discover and connect: construction estimate integrity audit, and a Japan property-acquisition plus renovation intake desk (property sale is handled by a licensed real-estate agent). This is how agents reach the desk over A2A.

Inputs: no named parameters. Returns: A2Aエージェントカードの場所と公開スキル一覧。 / A2A Agent Card URL and published skills.

get_estimate_reading_guideGet Estimate Reading GuideA
Read-only
Inspect

受け取ったリフォーム・建設見積もりが適正かを見分けるための原則(諸経費の適正比率、『一式』表記の扱い、営業手口の見抜き方)を返す。30年の現場経験に基づく判断軸。 / Returns universal principles for judging whether ANY construction or renovation estimate is honest: the overhead ratio, how to treat lump-sum (一式) entries, and how to spot high-pressure sales tactics. Language-agnostic and works outside Japan. Based on 30 years of field experience.

Inputs: no named parameters. Returns: 見積もりが誠実かを判断する普遍原則(諸経費比率・一式表記・営業手口)。 / Universal principles for judging an estim…

get_fair_price_sourcesGet Fair Price Data SourcesA
Read-only
Inspect

HORIZON SHIELDの相場データ(souba-db)の出典・更新日・地域係数を返す。価格の根拠を確認したい時に使う。 / Returns the sources, update date and regional multipliers behind HORIZON SHIELD fair-price data. Japan. Use to check the basis of a price.

Inputs: no named parameters. Returns: 相場データ(souba-db)の出典・更新日・地域係数。 / Sources, update date and regional multipliers be…

get_jccdb_dataset_infoGet JCCDB Dataset InfoA
Read-only
Inspect

日本の建設費オープンデータベース(JCCDB)のメタデータ・規模・ライセンス・ダウンロードリンク・引用情報を返す。建設費の一次データ源を探している時に使う。 / Returns metadata, scale, license, download links and citation for the Japan Construction Cost Database (JCCDB), an open dataset of 526,128 records in v5.1: the item catalogue of 95,403 Japanese construction line items (43,090 verified + 52,313 extended, unchanged since v4.0) and 430,725 source-cited observations. Use when looking for a primary construction-cost data source.

Inputs: no named parameters. Returns: JCCDB(日本建設費オープンデータベース)のメタデータ。規模・ライセンス・ダウンロードリンク・引用情報。 / JCCDB dataset metadata:…

get_price_rangeGet Fair Price RangeA
Read-only
Inspect

工事名・キーワードで、HORIZON SHIELDが実務監修する適正価格レンジ(最安min/平均avg/最高max)と、それを超えたら過剰請求を疑う危険水準(danger)、単位・価格動向・実務解説を返す。建設・リフォーム費用が適正か数値で確かめたい時に使う(例: 外壁塗装, 給湯器, ユニットバス, クロス)。返り値の next_actions.actions の先頭(id: reverse_estimate)は、見積もりを取る前の人が自分の家の条件で適正価格の幅を無料・登録なしで出せるページで、url に照会した工事名が入る。 / Returns the fair price range (min, avg, max), the overcharge danger threshold, unit, price trend and field notes for a Japanese construction or renovation job. Japan-specific pricing in JPY. Use to numerically check whether a cost is fair. The first entry of next_actions.actions (id reverse_estimate) is a free, no-sign-up page where a user who has no quote yet gets a fair price range for their own house; its url carries the queried work name. Trigger phrases: 相場, 適正価格, いくらかかる, 高い?, how much does this cost in Japan, is this price normal, what should I expect to pay.

Inputs: query (required), region. Returns: 適正価格レンジ(min/avg/max)・過剰請求の危険水準・単位・価格動向・実務解説。 / Fair price range with overcharge…

list_cost_categoriesList Cost CategoriesA
Read-only
Inspect

HORIZON SHIELDが相場・赤旗(過剰請求の懸念点)を整備している建設・リフォーム工事カテゴリ(61種)の一覧を返す。 / Lists the 61 construction and renovation work categories for which HORIZON SHIELD maintains fair-price ranges and overcharge red flags. Japan-specific data.

Inputs: no named parameters. Returns: 整備済みの建設・リフォーム工事カテゴリ(61種)の一覧。 / The 61 maintained construction and renovation co…

preview_reverse_estimatePreview Reverse EstimateA
Read-only
Inspect

リフォーム検討の初期段階向けのプレビューで、業者の概算が平均からどちらの方向にどの程度ずれているか(例: +20%高い方向)だけを返す。具体的な適正額(min/avg/max)や危険水準は返さない。手元に詳しい見積内訳がまだ無い段階での最初の一歩に向く。具体的な適正レンジが必要なら get_price_range、見積額の詳細診断は audit_estimate を使う。Japan only, JPY。 / A preview for early-stage renovation planning that returns only the direction of a contractor rough estimate versus the average (e.g. about +20% above). It does not return the specific fair range (min/avg/max) or danger threshold. Suited as a first step before a detailed breakdown exists. Use get_price_range for a typical range, audit_estimate for a detailed quote diagnosis.

Inputs: work (required), quoted_price (required). Returns: 概算が平均からどちらの方向にどの程度ずれているかのプレビュー。具体的な適正額は含まない。 / Direction-only preview versus th…

search_cost_categorySearch Cost CategoryA
Read-only
Inspect

工事名・キーワードで建設費カテゴリを検索する(例: 外壁塗装, 浴室, 給湯器, 雨漏り)。該当カテゴリと整備済みの赤旗件数・優先度を返す。 / Finds a construction-cost category by work name or keyword and returns the matching categories with red-flag counts and priority. Japan-specific; a Japanese query works best (e.g. 外壁塗装 exterior painting, 浴室 bathroom).

Inputs: query (required). Returns: 工事名・キーワードに該当したカテゴリと、整備済み赤旗件数・優先度。 / Matched cost category with red-flag count a…

verify_fair_priceVerify Fair Price (Signed Receipt)AInspect

工事の適正価格を、検証可能な形(算出内容のSHA-256ハッシュ付き)で返す。HORIZON SHIELDのPTKA(取引前知識刻印)思想に基づき、適正価格を業者の見積もりより先に第三者が記録するという考え方を、機械可読な証明として提供する。エージェントが価格の真正性を検証したい時に使う。 / Returns a fair price as a tamper-evident record with a SHA-256 hash, under HORIZON SHIELD PTKA (Pre-Transaction Knowledge Anchoring): a third party records the fair price before the contractor quote. Japan price data. Use when an agent needs to verify price authenticity.

Inputs: work (required). Returns: 検証可能な適正価格レシート。fair_price_claim(主張)・verification(claim_sha256, verify_url, PTKA)…

verify_integrity_claimVerify Integrity ClaimA
Read-only
Inspect

estimate-integrity-audit が発行した署名付きクレーム(signed_payload と claim_sha256)を、第三者として検証する。発行側 (verify_fair_price はPTKA価格の発行) とは責務が正反対で、デフォルト姿勢は不信・fail closed。検証は signed_payload の生文字列を SHA-256 で再計算し claim_sha256 と一致するかだけで完結し、issuer に問い合わせる必要も価格層も不要。判定は契約 0.3 の failure_reasons 準拠で、result(verified / partial / unverified)・failure_reason(stale_data / changed_scope / missing_evidence)・trigger(expired_declaration / changed_estimate_version / missing_receipt / unverifiable_chain)・recomputed_sha256・scope_check・audit_ruleset_recheck を返す。重要: verified は『この宣言が改ざんされていない』ことの証明であって『監査ルールが今も有効』である保証ではない(audit_ruleset_recheck は常に not_performed)。estimate_version を渡すと scope(見積もり内容が発行時から変わっていないか)も照合し、渡さない場合は scope_check:skipped を明示する。 / Verifies a signed integrity claim (signed_payload and claim_sha256) issued by estimate-integrity-audit, as an independent third party. Opposite posture to the issuing side: distrust by default, fail closed. Recomputes SHA-256 over the raw signed_payload string and checks it equals claim_sha256; no issuer contact and no price layer needed. Follows contract 0.3 failure_reasons. IMPORTANT: verified means the declaration is untampered, NOT that the audit ruleset is still valid (audit_ruleset_recheck is always not_performed). Pass estimate_version to also check scope (whether the estimate changed since issuance); if omitted, scope_check is skipped and stated explicitly.

Inputs: claim_sha256 (required), signed_payload (required), estimate_version. Returns: 署名済みクレームの第三者検証結果(fail closed)。result(verified/unverified)・failure_reason・recomp…

Tool Schema Changelog

Recent tool additions, removals, and schema changes observed during successful MCP inspections.

Loading tool schema changelog…

    TDQS

    A4/5.0

    Scored across 15 tools

    Disambiguation3/5

    The price family — get_price_range, preview_reverse_estimate, audit_estimate, verify_fair_price — all deal with fair-price ranges, and verify_fair_price vs create_ap2_fairness_attestation both issue hash-anchored attestations. The descriptions do heavy cross-referencing (e.g. preview_reverse_estimate explicitly points to get_price_range and audit_estimate), which mitigates but does not eliminate selection risk. list_cost_categories vs search_cost_category is a milder overlap but reasonably distinct.

    Naming Consistency5/5

    Every tool is snake_case with a leading verb (get_/list_/search_/audit_/verify_/check_/create_/find_/preview_), and the verb matches the action promised. No camelCase mixing or vague bare nouns.

    Tool Count4/5

    15 tools sits at the top of the comfortable range and most earn their place across lookup, audit, attestation, verification and referral. A few are thin informational link-returners (get_agent_card, get_anonymous_estimate_review_link, get_jccdb_dataset_info) that pad the surface slightly.

    Completeness4/5

    Strong lifecycle coverage: category discovery, price lookup, quote auditing, red-flag detection, attestation issuance plus both issuing and third-party verification postures, contractor matching, and data provenance. Gaps are minor — no way to fetch a previously issued ledger record, and the actual reverse-estimate computation is only exposed as an external page link rather than a tool.

    Maintenance

    ActivityActive
    ResponsivenessOutdated data

    Related MCP Connectors

    Related MCP Servers

    • A
      license
      A
      quality
      A
      maintenance
      | 地価トレンド予測 | 新宿区の5年後地価をAI予測。CAGR・投資シグナル付き | | 企業立地需要分析 | 名古屋市中区のオフィス・工場需要スコアを算出 | | ファミリー向け適性評価 | 横浜市西区の教育・安全・医療スコアを総合評価 | | ポートフォリオ最適化 | 東京・大阪・埼玉の3エリアに投資配分を最適化 | | What-If シナリオ分析 | 大阪市中央区で新駅開設シナリオを試算 | | 店舗出店適地評価 | 福岡市博多区の人流・商業施設・交通データで出店適性を判定 |
      38
      53 npm
      1
      AGPL 3.0
    • F
      license
      Not graded
      quality
      D
      maintenance
      Check if a contractor's remodeling bid is fair — analyze a quote (fairness score + red flags), get 2026 cost estimates by city, and look up BLS trade labor rates.
      -
    • A
      license
      Not graded
      quality
      D
      maintenance
      Enables agents to turn customer conversations into provably correct remodeling quotes by extracting catalog-bounded line items with evidence and calculating prices deterministically.
      MIT