Receipt · a demo by Geeq

How do you know why version 27 was accepted?

Not what changed. Why it was trusted.

See the problem
01Present day

A normal pull request, on a normal Tuesday.

Alice tightens a safety threshold in CLAUDE.md. Reviewers approve, checks go green, she merges.

Opengeeq/agent-policies · #274alice/tighten-eval → main

Update CLAUDE.md: tighten eval threshold for prompt safety checks

opened by Alice Chen

  • Alice Chen approved
    1 reviewer • required
  • CI passed
    ci / build & test • 2m 41s
  • Security scan passed
    codeql • 0 findings
  • Prompt eval: 0.94 (threshold: 0.90) — passed
    eval-harness v3.2.0
All checks passed
02Six months later

Fast-forward six months.

Nothing dramatic. Just the ordinary drift of a real system.

Mar 14, 2026+ 6 months
||||
  • Alice Chen left the company
    offboarded · Aug 2026
  • The evaluation model has evolved (v3 → v4)
    eval-harness release
  • A dependency now has a published CVE
    CVE-2026-31840 · high
  • The repository has continued to change
    312 commits since merge

drag the slider →

03The gap
a3f9e21 — Update CLAUDE.md · merged by Alice Chen · Mar 14, 2026

Git tells you what changed.
It doesn't tell you why it was trusted — or whether you still should.

  • Which eval model scored this?
    ✕ unknown
  • Was 0.94 actually above threshold at the time?
    ✕ unknown
  • Who's accountable if this breaks in production?
    ✕ unknown
04The fix

What if the approval itself was the evidence?

Not a log you go find later. A signed receipt, emitted at the moment of decision.

Attestation Receipt
Verifiable

CLAUDE.md eval threshold change approved

Approver
Alice Chened25519:9f4c…a17b
Timestamp
2026-03-14T16:42:18Z
Attestation hash
0x7d3a…f19c8b
Evidence bundle
  • ci_rungh-actions/run/8821447301
  • sec_scancodeql/scan/44e2a1
  • eval_modeleval-harness v3
  • threshold0.90
  • eval_score0.94

Signed and notarized at the moment the decision was made.
Anchored on Geeq.

The receipt permanently preserves the conditions and reasoning behind the decision, so it can be verified — byte for byte — at any point in the future.

05Verify

Six months later…

  • · Alice has left the company.
  • · The evaluation model has evolved.
  • · A dependency now has a published CVE.
  • · The repository has continued to change.

Despite all of that, the original approval can still be verified — because the evidence was captured when the decision was made.

  • CI rungh-actions/run/8821447301
  • Security scancodeql/scan/44e2a1
  • Eval modeleval-harness v3
  • Threshold0.90
  • Eval score0.94
  • Approver signatureed25519:9f4c…a17b
Decision Evidence
geeq/agent-policies#274
0x7d3a…f19c8b
no live pipeline · no personnel · just the receipt
06How it works

The SDK doesn't replace anything you use. It just remembers why.

01

Your existing tools run as normal

GitHub, CI, eval framework. No new pipeline.

02

A few lines of SDK emit a signed receipt

When a decision is made, the evidence is bundled and signed.

03

Anyone verifies it, later

A person, or another agent. No rebuild, no archaeology.

import { receipt } from "@geeq/receipt";

await receipt.attest({
  decision: "merge-pr",
  subject: "geeq/agent-policies#274",
  evidence: { ci, security, eval },
  approvedBy: user.id,
});

Try it on your own CLAUDE.md.

Geeq IGNITE — bring your own repo, we'll show you the receipt.

·Receipt · a demo by Geeq