# Case record: one morning's work, September 22, 2026

Orbyt Labs, R&D evidence, version 1.0 (snapshot October 20, 2026). Times are UTC, read from the repository's
commits and run records. Run ids, commit hashes and session transcripts are withheld from this public copy and are
available in diligence.

## Why this case

It is one real sequence that crossed company functions (operations, engineering, quality), was caught by an
automated check, was repaired by a founder-launched headless agent with no prompt from him in between, and was then shown fixed by the same seat
running again. It was chosen because its records are complete, not because it is typical. A single case shows that
the mechanism worked once; it does not show how often it works.

## The sequence

| Time (UTC) | Function | What happened | Record |
|---|---|---|---|
| 04:37 | Operations | The rollout ladder (a deterministic evaluator, no model) committed its daily state. | Bot commit by the ladder workflow |
| about 04:38 | Engineering | The engineering seat was dispatched and began its health audit (84 agent turns). | Seat run row, trigger `workflow_dispatch` |
| before 05:28 | Quality | The workflow's gate refused the seat's output. The ladder's commit had changed state that a public status file is derived from; the file on main was stale, and the site's tests caught it. The run was recorded as `rejected`, reason `guard-refused`, beside the agent's claim that it had finished. | Seat run row (landed 05:27) |
| 05:43 | Engineering | A headless handoff-loop iteration (Claude Code, launched earlier by the founder with a written queue; no prompt from him during the diagnosis and repair) replayed the failure, changed the seats' shared housekeeping configuration so every seat rebuilds the derived public files before its checks, added a test that fails without the change, and landed the fix. Continuous integration passed on it. | Commit and CI run |
| 06:10 to 06:43 | Engineering | The engineering seat was dispatched again on the fixed tree; its report landed at 06:43 (45 turns, completion). | Seat run row, bot commit |
| 07:15 | Operations | The deterministic daily pulse recorded September 22's state; its closer closed the last open automation alarm. | Bot commit, issue closure |
| September 27 | Engineering | The next scheduled engineering run completed. | Seat run row, trigger `schedule` |

## The human part

- The founder launched the loop and wrote the queue it worked from. The same night he committed unrelated site fixes.
- The repair session's transcript shows the agent's edits, a failing control before the fix and the passing tests
  after it, with no human prompt between the failure and the fix.
- After the fix, the loop's own wrapper stopped on a separate defect (a dirty working tree), so the iteration as a
  whole did not complete cleanly.

## The mechanism that made the refusal legible

On September 14 a change landed that records a guard's refusal beside the agent's claim of completion (the agent
says it finished; the gate says it did not; both are kept). Before it, refused runs carried no reason. This
morning's refusal is one of the three production refusals it has named since.

## What this case does not show

- How often such a repair succeeds, or how often a refusal has no clear cause.
- How much founder time the system needs over weeks. The founder's time is reconstructed only as a floor and is
  almost entirely unclassified by purpose.
- Whether the repair would have been accepted by an independent reviewer; continuous integration and the seat's
  rerun are the evidence offered.

## Verifying it in diligence

The run ids, commit hashes, the CI run, the issue numbers and the repair transcript lines can be shown from the
private repository on request.