Release notes — v1.0.1

MCP server: verify_claims tool

v1.0.0 shipped the MCP server with check_evidence only — the verify_claims
tool was still on main mid-merge when v1.0.0 was tagged and published. This
patch adds it, so the MCP server now exposes both halves of the proof loop:

• check_evidence — before generation, may the model speak at all?
• verify_claims (new) — after generation, does the answer stand on the
  evidence: every citation resolves to a real record, every claim carries a
  citation, and — when records carry facts — every cited number matches them
  exactly (misquoted_values).

Accepts { answer, records, supporting?, gate?, preset?, rules?, decision? }.
With the same decision.id and records as the check_evidence call, its
verification record joins that decision record on an identical evidence
digest — the audit trail is complete over MCP, not just the direct API.

No other changes. The Python package (evidence-gate-py) has no code changes
in this release — its version bump is for parity across the four version
points (npm, PyPI, __init__.py, MCP server), per this project's convention.

Install

npm install evidence-gate@1.0.1
pip install evidence-gate-py==1.0.1

Links

• Full changelog: CHANGELOG.md
• MCP usage: README

Evidence Gate 1.0 — the proof loop is complete

Evidence Gate 1.0 — the proof loop is complete

Evidence Gate is the missing step between retrieve and generate: a tiny,
zero-dependency gate that decides whether an LLM may speak from your data — and
now, whether what it said actually stands on that data.

This 1.0 closes the loop end-to-end:

• Gate (evidenceGate) — before generation, decide summarize / compare
  from evidence coverage, freshness, and quality; get prompt-ready caveats.
• Prove the answer (verifyClaims) — after generation, check every citation
  resolves to a real record (no phantom evidence), every claim carries a citation
  (no naked claims), and the framing matches the verdict (no "as of today" over
  stale data). Deterministic, no model call.
• Catch misquoted numbers (record.facts) — opt in by attaching
  facts: { revenue: 1234500 } to a record, and every number in a sentence
  citing it must match exactly, or the answer fails with misquoted_values. A
  correctly cited but wrong figure — the failure mode a citation check alone
  misses — is now caught deterministically. Magnitude suffixes (1.2M,
  1.5 ล้าน), thousands separators, and Thai numerals all normalize identically
  across ports.
• Prove the evidence (record.provenance) — where each observation came
  from, hash-linked transform chain, source authority; surfaces as caveats and a
  replay-verifiable digest.
• Audit trail — decision and verification records join on the request id and
  an identical evidence digest, without storing the evidence or the answer.
• MCP server — a check_evidence tool so any MCP-compatible agent can gate
  itself before it speaks.

Everything is byte-identical across the JavaScript and Python ports, locked by a
shared vector file (test/vectors.json) that both test suites run.

⚠️ Breaking change

Rulesets are now validated at call time: staleDays, minRecords, and
qualityThreshold are required finite numbers. A ruleset missing one throws
instead of silently behaving as permissive. Presets are unaffected — only
hand-built rulesets need to ensure these three fields are set.

Links
Full changelog: CHANGELOG.md
The proof loop, provenance, MCP usage: README

Install

npm install evidence-gate
pip install evidence-gate-py