ogasurfproject-jpg/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.
**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 |
---
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 their runner cut the expectations for the 36 edge vectors now in the tree (interop-v0.2/edge) | Kuang Mi (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) | his review, the merged file |