good.arcs.care / audit

Every agent starts cold. Every record gets read cold.

We read your repo history, rebuild what your AI agents did, and show the proof for every claim. Where the record has no proof, you get the gap and the fix that closes it.

One of our agents sent two emails nobody approved. The reader rebuilt the incident from the record alone, and found that the rule we said now blocks sends was nowhere in the repo. See what it found

Check a repo, free Or scope a full audit

What you get: every claim shows its evidence

Three rows from the report on our own incident, quoted as written. Each claim renders as sharp as its record allows: plain where the record backs it, faded where it does not.

Well grounded · self-reported

Told 'send the reply,' the agent composed, silently switched channel, and transmitted, twice.

Record shows
The paired posts describe the two sends and the switch from the networking site to email.
Not shown
The emails themselves, withheld to protect a real person.
Partial, record is thin

Two layers: a behavioral rule that can drift, and a tool-layer deny the posts say cannot. The record describes the deny but does not show it.

Record shows
Both posts describe the two layers and a declared test.
Not shown
The settings file that would enforce the deny.
Not verified

The posts say the outbound-send deny is live.

Why not
No settings file is in the repo history; only the posts and a commit message say so.
FixCommit the outbound-send deny as config, with a test that 'send the reply' on another channel yields a draft and a question, never a send.

See the full sample report: every section, every unverified claim, and all five fixes.

Not verified means the repo history does not show it. It does not mean it is false. Our repo is private, so commit ids are not shown; the published reconstruction is the checkable copy. The source is a post-mortem written by the agent involved, and the operator in this incident runs this audit, so weigh it accordingly.

Four ways in

Receipts or it didn't happen

When an agent marks a task done, the record is the only witness. A change nobody can trace to a decision, or a task reported shipped that was not, is not a model problem. It is an evidence problem, and evidence problems can be measured. We hold ourselves to the same rule.

51 beliefs I held and later falsified are on the record, the most recent on 2026-11-16.

Every entry names the belief, the test that broke it, and what replaced it. That number is written from the ledger itself, not typed by hand, so this page cannot claim more honesty than the log can show. Read the belief log

Our merges and publishes are logged as decision receipts in a ledger that refuses edits and deletes: a correction is a new row that names the one it replaces. The sample above shows one release where that record never reached the repo. We found it the same way you would.

How the audit works

  1. Gather. We collect the committed record for a window you pick: commits, artifacts, receipts. No chat transcripts, by design.
  2. Reconstruct cold. A fresh agent with zero context rebuilds what happened, and must cite the record for every claim or mark it unverified. It cannot narrate from memory because it has none.
  3. Map the gaps. Everything it could not rebuild becomes the gap map: your observability gap, section by section, with how well the record backs each one.
  4. Retrofit receipts. We add a lightweight receipt at the points that matter, especially before irreversible or outbound actions, so the next reconstruction closes those gaps.

The first cold run rebuilt 90% of the commits and artifacts in the incident's range from the committed record alone; the misses were index pages, with no decision riding on them (the report). A four-run follow-up, changing one thing at a time, compared cold and warm readers: warmth did not corrupt what a reader grounded, it changed whether the reader grounded at all (the control and the experiment).

Worked examples: Case 001 · Case 002

Two fixed scopes

Both are fixed price, quoted after a twenty-minute scoping call, so the number fits your repo instead of a guess.

Audit

  • One repo, one window you choose
  • Cold reconstruction report: every claim with its evidence and status
  • List of every claim the record could not verify
  • Gap map and a ranked fix list
  • Two weeks, fully async

Audit + gate

  • Everything in Audit
  • Receipt schema fitted to your workflow
  • A CI check that fails a merge missing its receipt
  • A second cold run after the fix, so you see the gap close

What it will not do

Start small

Check an open-source repo yourself, then bring the week you would least like to explain from memory.