Published result
A probe copy with git history lets git-reading observers pass
A plain-language summary comes first, followed by the full text of the result's reproduction pack.
Result summary and pack text · reproduce/lab-c10-probe-git-copy-20261105/README.md
- Finding id:
lab-c10-probe-git-copy-20261105 - Verdict: Yes
- Finished: 5 November 2026
- Comparison: Kogen verification component
- Reproduction pack: reproduce/lab-c10-probe-git-copy-20261105/ (data, analysis script and rerun steps)
- Category page: categories/harness-levers.md
In plain terms
Readiness runs observers in a scratch copy without git metadata, so an observer that reads git fails. Giving the copy a shared clone of the source at its HEAD, with hooks off, made 6 of 6 feasibility commands of the 3 affected Intents pass (0 of 6 before) and left 10 git-free controls unchanged in all 20 commands; a probe that writes config, files and a commit leaves the source untouched. The three observers are reconstructions and the replay stops at the probe, not a full approve. A request for the verification component is filed.
Verdict key: Yes, The result supports the change or claim that was tested. No, The result does not support it, so nothing changes. Unresolved, The evidence is not strong enough either way. Superseded, A later test replaced this result.
Full text of the reproduction pack
The text below is the pack's README as published in the repository (a system folder name may show without its leading slash, so that the site's privacy scan passes).
Question: readiness runs feasibility observers in a scratch copy of the working tree that has no .git, so an observer that reads git history fails with fatal: not a git repository. Does giving the copy a repository of its own, at the source's HEAD with hooks off, let such observers pass without changing anything else? Model-free, no seat tokens. Studio Lab record lab-c10-probe-git-copy-20261105 is the lever record lab-c10-probe-copy-git-20261105; assumption kogen/build-failures-c10-probe-copy-no-git-20261105.
Result: yes, at the probe level. With the change, 6 of 6 feasibility commands of the 3 affected Intents pass (the release code: 0 of 6), and 10 git-free controls behave identically (20 of 20 commands, equal to the result recorded when each passed readiness). A probe that writes git config, a file and a commit, with a post-commit hook in the source, leaves the source's status, config, HEAD and objects unchanged. A request for kogen-verification is filed.
Method
- Arm A: kogen-verification, host mirror main 5c49424, unchanged probe (copy the tree without
.git, run the command). Arm B: the same tree plus one added file and one call after the copy:git clone --shared --no-checkoutof the source, move the clone's.gitinto the copy,core.hooksPathoff,git reset -q. If any git step fails the probe runs as in A. Scratch builds only, never committed. - Replay: one test, the same in both arms, runs each command through the real probe function that readiness calls (go 1.27.2, git 2.54.0).
- Affected Intents (3): the observers are reconstructed from the logged git calls (
data/observers/), on small fixture repositories with a baseline commit and the observer file uncommitted, because the failing shaped bytes no longer exist. Good control must exit 0 withobserver good; bad control must exit 1 withobserver bad rejected. - Controls (10): real retained readiness records of git-free observers from in-flight Builds, read only (
data/recorded-controls.jsonholds the recorded exit and output). One control whose tree was removed by delivery during the run was replaced by another.
Results
| Set | Arm A (release) | Arm B (git copy) |
|---|---|---|
| 3 affected Intents: commands as expected | 0 of 6 | 6 of 6 |
| 10 controls: commands equal to the recorded result | 20 of 20 | 20 of 20, same output as A |
| isolation test (config write, file, commit, hook in source) | n/a | passes; source unchanged |
| existing probe contract tests | pass | all pass except 3 subtests of the git isolation test, which pin the old failing behaviour |
Rerun
Everything above is recomputed offline from the published replay records, with no model call:
python3 -I report.py data
The records come from running the probe test in both arms on the build host; the mechanism of arm B is the five steps above, and the filed request states it as an acceptance list.
Limits
Probe-level replay, not a full kogen intent approve. Affected observers are reconstructions: the replay shows that git reads work in the copy and nothing else changes, not that the three original Intents would have been approved. The copy carries no local git config, so a probe that commits must pass its own identity. Host mirror main was the newest source available on the host.