freedom-intelligence/docs/evidence/2026-09-23-brief-defects.md

28 lines
1.6 KiB
Markdown
Raw Normal View History

# Compatibility executor defect log
Baseline for the FI-WP-0005-T07 canary comparison: defects produced by the
current `rein-aharness fi-research-brief` path (single llm-connect call, no
network access, no primary sources). The canary must not reproduce these.
| Date | Defect | Class | Evidence |
| --- | --- | --- | --- |
| 2026-08-04…14 | DeepSeek-V4-Flash-0731 reported as a 12B dense model; once claimed as already collected | invented facts; false collection claim | Corrected in `inventory/catalog/deepseek-ai__DeepSeek-V4-Flash-0731__strategic.yaml` history (FI-WP-0004-T05) |
| 2026-09-23 | "Qwen3.8-27B has been moved from 'candidate' to 'approved'" | invented state change; contradicts the catalog | Catalog unchanged since `a78187c`; `approved_by` empty. Correction appended to `briefs/2026/09/2026-09-23.md` |
## Pattern
Both defects are assertions about **lab state** that the brief had in its own
context and still got wrong, not stale knowledge about the outside world. The
executor receives the catalog and the previous brief, then narrates a change
that no file records.
Consequences for the FI-WP-0005 acceptance run (T07):
1. Compare canary output against the catalog for every claim about collection
status, sizes and licenses. Any assertion of a state change must cite the
file and revision that made it.
2. Briefs with no external delta should say so, rather than manufacturing an
internal one. Both defect days had genuinely quiet outside channels.
3. A false claim propagates: the previous brief is context for the next one.
Corrections belong in the brief file, not only in the workplan.