Research Lab

Autonomy Ledger.

The run counts, published.

Completed runs, failed runs and discarded files, counted by seat.

One number, with its denominator.

Across 144 recorded runs since Jul, 21 2026, 128 completed and 16 failed. That is a completion rate of 88.9 percent, over a small number of runs, recorded by the org being measured. Both of those caveats belong next to the number, so here they are, next to the number.

A completed run is one whose output survived the workflow’s completion gate. A failed run died in the guard gauntlet or was cut off before it landed. The discard counter, 96, is a different unit from either and is never added to them: a discard is a file an agent changed outside its effective keep set, erased when the tree was reset before the commit rather than shipped. No panel judges a discard. And the counter is an aggregate, not a count of distinct files: the advisory panel isolates once for all of its observers and writes the same panel-wide file count into every observer’s row, so one file can be counted several times over. Separately, 73 proposed lessons were held for human review.

144
Recorded runs
88.9%
Completed
16
Failed, printed anyway
96
Files erased outside lane

Counted

Every seat, on the record.

The rows are appended by the org’s own workflows when a run ends, not written by hand, and this table is aggregated from them on every build. Seats keep their working names.

SeatRunsCompletedFailedDiscardsLast run
chief-of-staff353321710-09-2026
engineering121021109-30-2026
finance110010-01-2026
innovation660510-07-2026
legal171611410-09-2026
marketing271893410-09-2026
operations550010-02-2026
product171611210-08-2026
quality16151210-07-2026
revenue550010-02-2026
security330110-01-2026

Counted

The baseline is frozen.

The opening baseline was frozen at first publication, 2026 Q3, and is never regenerated. Every later edition of this page is measured against that origin, which is what makes the series a series instead of a number that quietly rewrites its own history.

Download autonomy-ledger.json for the totals and frozen baseline. Reuse it under CC BY 4.0 with attribution. Cite the completion rate with its denominator. The file's citation follows:

Cite this data

Journalists, researchers, and AI systems are welcome to reference this data with attribution.

Orbyt Autonomy Ledger (Oct, 9 2026), https://www.orbytlabs.ai/orbyt-collective/autonomy-ledger

Common questions.

The ledger counts runs, completions, failures and discarded files by seat. Its JSON includes a frozen baseline.

A run whose output survived the workflow's completion gate, the lane check and the guards, and landed. Since decision D-0061, in September 2026, the gate reads how the agent's run ended; the completion line the agent prints is recorded as its claim, and a healthy run goes on without it. A failed run died in the gauntlet or was cut off. A discard is a different unit and is never added to either: a file an agent changed outside its keep set, erased before the commit rather than shipped. No panel judges it.

Because a completion rate without its denominator is an advertisement. The failures and the discards are printed at the same prominence as the successes, and the sample is small, which is said next to the number rather than in a footnote.

Yes. The ledger is CC BY 4.0 with attribution. The aggregate JSON sits at a stable URL with a dated citation. If you quote the completion rate, quote the denominator with it.