Local results, with their limits.
MacBook checks recorded on 16 November 2026. Published as documentary evidence, not a corporate-pilot acceptance.
Environment and source
RC8 source 6abc06f, Docker Desktop, Apple Silicon running Linux amd64 images under emulation. The cabinet used an owner token. Real model calls used an external OpenAI API through PR-Agent. Qwen was not tested.
The owner supplied a normalized report on 17 November. This publication checks and summarizes that record; it does not rerun the tests. The original private report and company source are not public downloads.
A real run. Educational code.
Recorded on a MacBook, 16 November 2026. A local GitLab and an external OpenAI model.
The recorded outcome
- SHA
8a87c4f9- Run
9f90f98e- State
draft_ready- Findings
- 3
The answer reached a saved draft
A real model reviewed a synthetic SQL-injection change. After a local credential-fingerprint patch, the run completed with three findings. The patch addressed a workflow conflict, not the SQL code.
This public record is a summary of the saved acceptance report. It does not contain the original MR, full finding text or cabinet footage. It is not a live demo or an independent rerun.
The patches are not an accepted release. Publication, fix, new-SHA review, CI and the team decision still need a complete acceptance run.
Results are separate experiments
| Check | Recorded result | Limit |
|---|---|---|
| Reviewer assignment | Coordinator polling started reviews; recorded on 6 MRs | Local GitLab; automatic publication disabled |
| Bot mention | One MR triggered by a mention | A small trigger check, not a reliability measurement |
| CI trigger | Pipeline assigned the bot; coordinator run reached draft_ready with 4 findings | A separate synthetic run; not the 3-finding case or an accepted blocking CI gate |
| Four historical MRs | Small changes reached draft_ready with no findings, one model call each | Unlabelled sample: process evidence, not proof of correctness or quality |
| Restart persistence | Saved run/draft records and Worker attempts matched; no repeated model calls | Not backup/restore, failure during a write or five concurrent reviews |
| Retrieval in shadow | A separate fixture prepared 5 support fragments, 6005 bytes; model received only the 474-byte diff | Retrieved fragments were not sent to the model; no proven review-quality benefit |
| Findings on real code | Worker returned two model findings; Control then failed with conflict | Defects not independently confirmed; no successful full rerun of the same MR after the patch |
Local patches are not the release
A credential fingerprint based on file device/inode metadata produced an observed conflict after a model response on Docker Desktop. A local patch removed those two fields. It needs regression coverage for replacement and revocation, inclusion in a fixed build and full package acceptance. The separate quota-message fix and installer/indexer workarounds are not presumed merged.
Installation required local network and filesystem workarounds, more Worker memory and a trusted local GitLab CA. The embedding timeout was changed locally under emulation. These observations are not copy-and-run production settings.
Before a corporate pilot
- Merge the reviewed patches into a fixed build and run credential replacement/revocation regression checks.
- Accept one complete MR, controlled publication, fix, new SHA, re-review, CI and human-decision path.
- Read a new useful model result through ContextForge, including access denial and revocation.
- Measure five overlapping real reviews, queueing, p95, limits, failures and full cost.
- Accept cabinet SSO, company and retrieval isolation, backup/restore and clean target Linux installation.
- Validate enforced retrieval on representative code and compatibility/quality of the proposed corporate Qwen route.
A finding, zero findings, a CI trigger, approval and merge are different outcomes. Quality, savings and full cost have not been measured.
A measurable evaluation
Start with one repository and a technical owner. Accept a fixed build, compare the same historical MRs with the current process, then agree a small live flow. Set success thresholds before starting. Track confirmed defects, false findings, misses, human handling time, waiting, all calls and retries, and operations. Expand to one more team only after accepted results and measured benefit.