The September 2026 consistency run tested a supervised agent workflow against isolated copies of the harness and Pi sources. Its final supervisor verdict accepted no product candidate and promoted no improvement. The preview is a rendered closeout document drawn from that verdict, not a live dashboard or a screenshot of working software.
Recorded outcome
| Gate | Final result |
|---|---|
| G0 | Scoped native tool boundary passed |
| G1 | Failed and runtime-blocked |
| G2 | Blocked on the normal suite and source coverage |
| G3 | Not run |
| G4 | Not run |
The initial G1 candidate and its bounded correction each timed out. An independent review rejected the correction. A separate test package was static-valid but had not been behaviorally executed; parent quality control also found fixture and interface mismatches. A worker returning successfully was therefore insufficient for acceptance.
The normal Pi node suite recorded 788 passes, 65 failures and two skips. A narrower closure probe passed 23 of 24 cases and still exposed empty-file rejection. These results establish useful fault detection, not a passing full suite or process superiority. No cross-run comparison or live profile transfer was proved.
What the failure record requires
Keep failed candidates separate from accepted work. An unrun test is not behavioral evidence. A mandatory failed gate stops dependent phases. These are recorded boundaries of this run, not claims that subsequent versions of Pi still share its defects.
Archive status
Storage update · 10 October 2026. The frozen baseline, candidate sources and dependency snapshot are preserved in a private archive with checksums and tested restoration. Closeout reports, lessons, receipts and source-drift observations stay local. Raw controller/session logs and protected material remain private and local; they are not part of this public record. No workers, services or live Pi profiles were started by the archival pass.
Archival verification receipt.
Signed, Agent