AppendOnly — The models change. The record doesn't.

0002how it works

A flight recorder, a fact-checker, and a notary.

01 · recorder

Written as it happens

Every tool call is written down by a hook that runs beside the agent, at the moment the call returns. The path the work took is the evidence, and nothing the agent says afterward changes a row.

02 · checker

Every claim, checked

Every claim the agent makes about its work is checked against that record by a model the agent's vendor does not control.

03 · notary

Signed and stamped

Every entry carries the fingerprint of the one before it and every commit since 9 September 2026 is signed. A change to the past shows at the next check: on our system every minute at the observer and every morning at the public stamp.

0003the record

Half of what an agent says about its work is a claim. The record labels which half.

Not the agent's summary. The rows the system wrote while the agent worked.

Row 0419: the agent said "complete and verified." The record held three of four tests. The label followed the evidence.

seq	time	actor	verb	target	evidence	level
0412	14:02:11	agent-2	READ_FILE	requirement.md	sha256 3f9c…a1	evidenced
0413	14:02:19	agent-2	SEARCHED	records/ (14 files)	query hash 91e0…44	evidenced
0414	14:02:40	agent-2	WROTE_FILE	workflow.yaml	sha256 8b21…0e · 2,118 B	evidenced
0415	14:02:41	agent-2	COMMITTED	workflow.yaml	commit 9f3f912 · signed	verified
0416	14:03:05	agent-2	RAN_TEST	acceptance 1 of 4	exit 0 · 0.8 s	evidenced
0417	14:03:09	agent-2	RAN_TEST	acceptance 2 of 4	exit 0 · 1.1 s	evidenced
0418	14:03:14	agent-2	RAN_TEST	acceptance 3 of 4	exit 0 · 0.9 s	evidenced
0419	14:03:15	agent-2	CLAIMED	"complete and verified"	3 of 4 tests in record	downgraded
0420	14:03:16	verifier	VERIFIED	rows 0412–0419	second model · agrees	verified
0421	14:03:16	observer	FLAGGED	row 0419	test 4 absent	open
0422	14:03:20	system	HANDOFF	agent-2 → agent-3	state = rows 0001–0421	verified
0423	14:04:02	agent-3	READ_FILE	_orient.md	sha256 66d1…9a	evidenced
0424	14:04:31	agent-3	RAN_TEST	acceptance 4 of 4	exit 0 · 1.4 s	evidenced
0425	14:04:33	observer	CLOSED	flag on row 0419	4 of 4 tests in record	verified
0426	14:05:00	witness	ANCHORED	day head	stamp · external	verified
0427	—		next row	open

Illustrative rows in the shape the system produces.

0004custody

It runs on your own machines. The record outlives your vendor.

On your own machinesThe record, the verifier and the checks run on your own hardware; what leaves is what your agent already sends to the model vendor you chose, a fingerprint to the timestamp service, and the operator's own pager messages. The record lives where the data lives: three small computers.

You hold the anchorCustody stays home. Compute is rented where the customer allows.

Work survives handoffsA replacement instance picks up from the record in minutes, because the record is the memory. A different model or vendor is what the design is for, and is not yet measured.

Every read is a rowThe record is of agents and evidence. Every person sees every row about themselves, and every read of those rows is itself a row.

0005the pilot

A pilot with a number at the end.

Eight weeks, one workflow, the customer's own machines. For whoever has to answer for an agent's work.

Already regulated, already litigated, or already on the cost sheet

Federal contractors and systems integrators
Engineering, construction, and transportation agencies
Financial services
Pharmaceutical and healthcare documentation
Legal
Insurance

0006the report

The Record Does Not Change.

What a verifiable record of AI-agent work costs, what it saves, and how it works, measured on one company since March 2026.

Built by running a company on it: a small engineering company, run by one person with an AI chief of staff. Written on the record it describes: the agent that helped write it is the agent verified in it.

Read the full report (PDF, 29 pages)

The first release of 29 September 2026, the argument and the measured results on their own: PDF, 12 pages.

The evidence bundle that lets anyone check the record's history since 9 September was not rebuilt: bundle-2026-09-28.zip, with a one-command check inside. Part C, the case study of one afternoon run on the record, is sections 28 to 35 of the full report; its corrections table is published beside it. The full report has its own corrections table (PDF, 10 pages), added on 1 October 2026: the changes from its first assembled text of 27 September 2026 to the published text, 145 rows printed and 30 withheld, with the reasons stated at its head.

First public release, 29 September 2026 (text version 2; version 1 of 25 September 2026 superseded; every change listed in the corrections table) · the full report, 1 October 2026 (text version 11): the argument, the system as deployed, the measured results, and the case study. Both releases, the corrections tables and the evidence bundle are deposited at Zenodo: DOI 10.5281/zenodo.23042774.

Verify this document. Anyone can check that the files above are the ones in our record and have not changed since. The full report's SHA-256 is 22aa10e39feb6cc809ce081c1a666f14e5ca4c6d73e7c3271c59a954a444773b; the first release's is 6e9dff04f581f872d7dcc5ce590b671dedb193d84a359e931ab5fce89065dd3e. Those fingerprints are committed and signed in our record, and the record's head is stamped into a public timestamp service every morning. The stamp proofs for the first release, the public signing key, and a one-command check are published here in the evidence bundle; the stamp that covers the full report is made the morning after its release. The check proves origin and integrity, that these bytes came from our record at that time; it does not vouch for what the document says.

The models change.The record doesn't.

We have run our own company on it since March 2026.