A code-repair workbench that keeps original source, proposed patches, independent checks, and observed outcomes inspectable together.
Engineering prototype
Recorded overview of a development investigation. The full investigation is linked in the case study; fresh validation is a separate result.Open full capture
Solo project; AI-assisted development.
Backend checks passed
428
14 October 2026 local release acceptance: 6 explicit opt-in skips, zero failures or errors. Software validation is separate from repair quality.
Eight evidence-viewer, eleven fault-fixture, four real-OIDC and one separate real-model development journey. These are different test scopes, not 24 repair successes.
12 October 2026 finite local HTTP/PostgreSQL workload over approximately 42 minutes, with zero unexpected errors. Fixture identity/tasks; no inference throughput or production-scale claim.
Ten-client, 30-minute soak within the same local workload. Successful-response percentile; zero non-success responses in this repeat. Earlier failed and interrupted runs remain recorded.
A plausible generated patch needs an inspectable source context, execution history, and independent checks before a reviewer can assess it.
The implementation
What I built
Built durable, inspectable repair workflows with fenced workers and independent verification. Local release acceptance passed 428 backend checks; model repair quality is evaluated separately.
Next.js
FastAPI
PostgreSQL
Python
Docker
OIDC
Features
Inspect original source and proposed patches
Review independent checks and saved evidence
Resume durable authenticated investigations
Architecture
Next.js evidence viewer and authenticated workbench.
FastAPI with PostgreSQL request receipts, ordered steps, tenant roles, and worker leases.
Local inference, disposable Docker checks, and digest-addressed artifacts.
Engineering decisions
Commit accepted requests before acknowledging them.
Keep interrupted inference explicitly unknown instead of silently repeating it.
Use independent checks to determine patch outcomes.
Validation and measured results
14 October 2026 local acceptance collected 434 backend tests: 428 passed, 6 explicit opt-in skips, zero failures or errors. Saved receipts: evidence/docs008/checks/backend-summary.json and fullstack-acceptance.json.
Twenty-four browser journeys passed across separate scopes: 8 evidence-viewer, 11 fault-fixture, 4 real-OIDC, and 1 exposed development-case model journey. This is software workflow coverage, not a repair-success count.
A saved Django development repair was reproduced against its registered checks. Fresh repair validation verified 0/10 attempts across five qualified cases; the development success is separate.
The completed 12 October 2026 local control-plane repeat processed 435,270 requests over approximately 42 minutes with zero unexpected errors. Its ten-client, 30-minute soak recorded acknowledgement p95 275.19 ms and read p95 16.79 ms. It used real HTTP/PostgreSQL and fixture identity/tasks, without inference or OIDC throughput. Saved receipt: evidence/build004/capacity-summary-final.json; earlier failures remain preserved.
The source repository is private. These notes summarize locally inspected receipts; source files and downloadable receipts have not been published here.
Limits and unfinished work
None of the ten fresh repair attempts across five qualified cases passed the registered verification checks. This measures model repair quality, not backend reliability. The single reproduced development repair was a separate, already-exposed case; broader repair usefulness remains unproven.
Recorded inspection does not run a new investigation or change a visitor's repository.
Local control-plane capacity does not establish production scale or inference throughput.
Saved release acceptance dated 14 October 2026 and capacity repeat dated 12 October 2026. Receipt filenames and scope are listed above; the repository is private.
