Every AI vendor asks to be trusted. This is the record that makes the question answerable instead: what our own AI did, taken from the same hash-chained ledger that /proof/ anchors publicly, generated by custody rather than written by hand.
It is small right now, and that is the honest state of it. It began on 2026-08-08. The point of a record like this is not how impressive it looks on day one — it is that it accumulates, that the uncomfortable rows are in it, and that nothing can be removed later without the chain showing it.
What our AI did — Automated Workflow
Every action taken by an automated model, what it was allowed to read, who signed it off, and how it turned out. Generated 2026-08-24T12:53:27 from a signed, hash-chained ledger.
25runs recorded
0stopped before running
25never approved
3produced nothing
0later found wrong
The uncomfortable part, which is the point. No accuracy rate is shown. 0 run(s) have been checked against a condition set for them in advance, and a percentage over fewer than 10 outcomes is noise. Counts are given instead.
25 run(s) were never approved by anyone. That is recorded rather than left blank, because a blank invites a charitable reading.
3 run(s) finished without producing anything. A job that reports success and produces nothing is the failure this whole system exists to catch.
25 run(s) were never approved by anyone. That is recorded rather than left blank, because a blank invites a charitable reading.
3 run(s) finished without producing anything. A job that reports success and produces nothing is the failure this whole system exists to catch.
Stopped before the model ran
These never reached the model, because the data it would have read was missing or out of date. A refusal is the most valuable line in this report: it is the wrong answer that was never written.
| when | agent | why it was stopped |
| No run has been stopped. That is only good news if the gate is switched on — an agent that declares no inputs can never be refused. | ||
Turned out wrong
Judged against a condition fixed before the outcome was known, so it could not be reinterpreted afterwards.
| when | agent | evidence |
| None recorded. Note what that does and does not mean: 25 run(s) were never checked against anything. | ||
Every run
| when | agent | model | could read | produced | approved by | outcome |
| 2026-08-24 06:43 | research-scout | claude-sonnet-5 | none declared | b38316e40ff9 | not approved | not checked |
| 2026-08-23 06:43 | research-scout | claude-sonnet-5 | none declared | 7588e98612c9 | not approved | not checked |
| 2026-08-22 06:43 | research-scout | claude-sonnet-5 | none declared | b30a39a510c6 | not approved | not checked |
| 2026-08-21 06:43 | research-scout | claude-sonnet-5 | none declared | df0fa6a85976 | not approved | not checked |
| 2026-08-20 06:43 | research-scout | claude-sonnet-5 | none declared | 06e7dadad54c | not approved | not checked |
| 2026-08-17 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | bd969da9e8c3 | not approved | not checked |
| 2026-08-16 06:43 | research-scout | claude-sonnet-5 | none declared | dc804856b726 | not approved | not checked |
| 2026-08-16 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | 168766feefe7 | not approved | not checked |
| 2026-08-15 06:43 | research-scout | claude-sonnet-5 | none declared | ea9b9ab60497 | not approved | not checked |
| 2026-08-15 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | b959257f6e6f | not approved | not checked |
| 2026-08-14 06:43 | research-scout | claude-sonnet-5 | none declared | produced nothing | not approved | not checked |
| 2026-08-14 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | 8c8265f34327 | not approved | not checked |
| 2026-08-13 06:43 | research-scout | claude-sonnet-5 | none declared | 8a996736cbb1 | not approved | not checked |
| 2026-08-13 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | 9bb0ccbca983 | not approved | not checked |
| 2026-08-12 06:43 | research-scout | claude-sonnet-5 | none declared | produced nothing | not approved | not checked |
| 2026-08-12 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | 4745dca44c85 | not approved | not checked |
| 2026-08-11 06:43 | research-scout | claude-sonnet-5 | none declared | 318a59280d7c | not approved | not checked |
| 2026-08-11 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | produced nothing | not approved | not checked |
| 2026-08-10 06:43 | research-scout | claude-sonnet-5 | none declared | de133d41fbec | not approved | not checked |
| 2026-08-10 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | 156110ba24a7 | not approved | not checked |
| 2026-08-09 06:43 | research-scout | claude-sonnet-5 | none declared | 177fcc95d031 | not approved | not checked |
| 2026-08-09 04:23 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | 77610d52c7fd | not approved | not checked |
| 2026-08-08 22:09 | portfolio-desk | claude-sonnet-5 | file #9fd5d540.txt | c4090df48edb | not approved | not checked |
| 2026-08-08 20:24 | research-scout | claude-sonnet-5 | none declared | c0b262290c2c | not approved | not checked |
| 2026-08-08 20:24 | wiring-test | claude-sonnet-5 | none declared | f873eef4f852 | not approved | not checked |
What this cannot prove. It records what a program declared and what changed
on disk. It shows which sources a run was entitled to read and that those bytes have not changed
since — not that the model read them, or read them correctly. It cannot see inside the model.
And it never claims an answer was true: only whether the condition set in advance was later met.
What it deliberately does not contain. No prompts and no outputs. The ledger keeps fingerprints and sizes, never the text.
Filenames are redacted. Sources appear as stable fingerprints, so this page can be handed to someone outside the company without showing them what the files are called — a filename is content too. The same source fingerprints the same way every run, so it is still possible to see that several runs read one source and another did not.
What it deliberately does not contain. No prompts and no outputs. The ledger keeps fingerprints and sizes, never the text.
Filenames are redacted. Sources appear as stable fingerprints, so this page can be handed to someone outside the company without showing them what the files are called — a filename is content too. The same source fingerprints the same way every run, so it is still possible to see that several runs read one source and another did not.