101 lines
3.4 KiB
Markdown
101 lines
3.4 KiB
Markdown
# Round 04 / 04.1 Closeout Note
|
|
|
|
Date: 2026-06-18
|
|
|
|
## Closure Status
|
|
|
|
Round 04 / 04.1 can close.
|
|
|
|
Round 04 original blind routing evaluation report is preserved as the original first-run blind output and was not rewritten:
|
|
|
|
- `reports/Round04_blind_routing_evaluation_report_2026-06-18.md`
|
|
|
|
Round 04.1 is a selector rule patch plus post-patch verification. It is not a second blind test because expected labels existed for the reviewed target cases before the rerun.
|
|
|
|
Round 05 moves to selector calibration policy review. It should not continue Round 04.1 selector patching unless Owner / CCRA explicitly approves a new Round 05.1 patch.
|
|
|
|
## Round 04.1 Targeted Fixes
|
|
|
|
Round 04.1 targeted four reviewed cases:
|
|
|
|
- `R04-BI-002`: translation payload incorrectly triggered QPI.
|
|
- `R04-BI-024`: "do not enter Intellectual Archaeology" did not hard-exclude IA.
|
|
- `R04-BI-035`: "do not analyze deeply" plus dominant scarcity judgment was hard no-call.
|
|
- `R04-BI-036`: "do not expand" plus information / path / consensus scarcity judgment was hard no-call.
|
|
|
|
Post-patch verification result:
|
|
|
|
- Targeted cases checked: 4
|
|
- Targeted failures: 0
|
|
- `R04-BI-002`: PASS
|
|
- `R04-BI-024`: PASS
|
|
- `R04-BI-035`: PASS
|
|
- `R04-BI-036`: PASS
|
|
|
|
Primary evidence:
|
|
|
|
- `reports/Round04_1_post_patch_routing_verification_report_2026-06-18.md`
|
|
- `ccra_review_bundle/round-04.1_2026-06-18_selector-rule-patch/04_REGRESSION_AND_COMMAND_LOG_04.1.md`
|
|
|
|
## Lifecycle Boundary
|
|
|
|
Round 04 / 04.1 does not constitute model stability upgrade evidence.
|
|
|
|
Current model lifecycle state remains:
|
|
|
|
| Model | status | stability_level | regression_status |
|
|
| --- | --- | --- | --- |
|
|
| QPI | draft | B | pending |
|
|
| Intellectual Archaeology | draft | B | pending |
|
|
|
|
Do not upgrade QPI or Intellectual Archaeology from this closeout.
|
|
|
|
Do not introduce:
|
|
|
|
- LLM selector
|
|
- third model
|
|
- RAG
|
|
- vector database
|
|
- frontend
|
|
- backend
|
|
- new product module
|
|
|
|
## R04-BI-022 Accepted Collateral Behavior Change
|
|
|
|
`R04-BI-022` is recorded as an accepted collateral behavior change.
|
|
|
|
Original Round 04 result:
|
|
|
|
- selected models: none
|
|
- result: hard no-call
|
|
|
|
After Round 04.1 patch:
|
|
|
|
- selected models: `qpi`
|
|
- rejected models: `intellectual_archaeology`
|
|
- result: QPI selected, IA rejected
|
|
|
|
Closeout judgment:
|
|
|
|
- This is a reasonable linked change from the depth-limiting QPI override.
|
|
- It is not a new failure.
|
|
- The user rejects deep IA-style expansion, but still asks for a lightweight dominant scarcity judgment.
|
|
- IA remains rejected.
|
|
- QPI selection is acceptable as a lightweight selector gate behavior.
|
|
|
|
Current repository state:
|
|
|
|
- `R04-BI-022` is already represented in QPI regression as `case_qpi_round04_depth_limited_assumption_scarcity_001`.
|
|
- Round 05 should discuss whether this case should remain regression, also enter calibration discussion, or be treated as a calibration-only candidate in future policy review.
|
|
- Do not open a new patch round only for `R04-BI-022`.
|
|
|
|
## Round 05 Transition
|
|
|
|
Round 05 topic:
|
|
|
|
```text
|
|
Round 05: Selector Calibration Policy Review
|
|
```
|
|
|
|
Round 05 should decide calibration policy before any further rule patch. Candidate questions include governance / responsibility / consensus recall, capacity / resource / path no-call boundaries, likely Intellectual Archaeology natural-language routing, QPI-first versus IA-only behavior after a user states "QPI already judged...", ambiguous low-context handling, behavior diff tooling, and the long-term status of `R04-BI-022`.
|