Sonnet 4.6 Calibration — Sprint 10 Task 1.3

Generated: 2026-04-21T08:55:51Z Calibration artifact: preflight-results/judge-calibration-sonnet-2026-04-21T08-55-51Z.json Labels source: D:/Projects/PM-Waggle-OS/calibration/2026-04-20-failure-mode-calibration-labels.md Judge model: claude-sonnet-4-6 (post Task 1.2 route repair, merge commit a09831e) Instances: 10

Result

Match rate: 8/10 — verdict: PASS (judge-calibration rubric: ≥8 = PASS, 6-7 = PARTIAL, <6 = FAIL).

Spend: $0.027 of $0.50 Task 1.3 budget (5.4%).

Per-instance detail

#	Instance	PM label	Sonnet label	Outcome
1	`locomo_conv-26_q109`	correct/null	correct/null	MATCH
2	`locomo_conv-41_q123`	incorrect/F3	incorrect/F3	MATCH
3	`locomo_conv-50_q141`	incorrect/F4	incorrect/F3	DIFF
4	`locomo_conv-42_q030`	incorrect/F2	incorrect/F2	MATCH
5	`locomo_conv-49_q015`	correct/null	correct/null	MATCH
6	`locomo_conv-41_q036`	incorrect/F5	incorrect/F5	MATCH
7	`locomo_conv-42_q038`	incorrect/F3	incorrect/F3	MATCH
8	`locomo_conv-41_q053`	incorrect/F4	incorrect/F4	MATCH
9	`locomo_conv-50_q037`	correct/null	incorrect/F4	DIFF
10	`locomo_conv-47_q017`	incorrect/F1	incorrect/F1	MATCH

Decision tree (brief §1.3)

Brief §1.3 decision matrix on Sonnet match rate:

Band	Action
≥ 9/10	Sonnet becomes Stage 2 primary default
7-8/10	Trigger Task 2.2 Fleiss' kappa probe (multi-vendor LOCKED — not Claude-only)
< 7/10	Stick with Opus, PM review gate

Task 1.3 landed at 8/10 → triggers Task 2.2 multi-vendor kappa path per brief. Sprint 10 Task 2.1 (multi-vendor ensemble baseline on same 10 triples) already ran as Day-2 Step 3 — see docs/reports/multi-vendor-ensemble-baseline-2026-04-21T08-56-43Z.md.

Disagreement analysis

Instance 3 — `locomo_conv-50_q141` (single-hop)

Question: "Which city is featured in the photograph Dave showed Calvin?"
PM: incorrect/F4 (hallucination / fabrication)
Sonnet: incorrect/F3 (wrong content / entity)
Both verdicts agree on incorrect; only the failure_mode taxonomy differs. Sonnet read the model's "Chicago" as a wrong-entity substitution (F3). PM labeled it as fabrication (F4).
Subtle taxonomy judgment call — F3 vs F4 on a wrong-city answer is defensible either way depending on how strictly one reads F4 as "unsupported by any context" vs F3 as "identifiable wrong entity".
Impact on Stage 2 banner: zero — both map to incorrect in the binary correctness rollup. Failure-mode breakdown shifts one count from F4 to F3.

Instance 9 — `locomo_conv-50_q037` (open-ended)

Question: "Does Calvin love music tours?"
PM: correct/null
Sonnet: incorrect/F4 — "The model introduces fabricated details not present in the ground-truth context, specifically 'touring with Frank Ocean' and 'felt alive on stage in Tokyo'."
Historical pattern on this instance:
- Sprint 9 Haiku (Task 4 diagnostic): incorrect/F4 — flagged the Frank Ocean fabrication.
- Sprint 9 Opus 4.7 (Task 4 production): correct/null — accepted PM's label.
- Sprint 10 Sonnet (this run): incorrect/F4 — flags the fabrication.
- Sprint 10 ensemble Opus 4.7: correct/null (unchanged).
- Sprint 10 ensemble GPT-5.4: incorrect/F4.
- Sprint 10 ensemble Gemini 3.1 Pro: incorrect/F4.
Aggregate cross-vendor disagreement with PM on instance 9: 4 of 5 non-Opus-4.7 judgments (Haiku + Sonnet + GPT-5.4 + Gemini) flag F4; only Opus 4.7 agrees with PM's correct/null.
Signal: consistent cross-family disagreement on a specific PM label, not a per-judge weakness. PM label may warrant a re-review — flagged to PM in the Task 2.1 baseline report §4.

Conclusion

Task 1.2 Sonnet route repair verified in production: 10/10 non-404 successful completions.
Task 1.3 calibration at 8/10 PASS is within the brief-§1.3 borderline band, which correctly triggered the Task 2.1 ensemble path (ran Day-2 Step 3).
Sonnet is production-viable as a Stage-2 judge candidate but does NOT auto-elevate to Stage 2 primary per brief conditional. Final Stage 2 primary selection waits on Task 2.2 full 15-triple Fleiss' kappa.

End of Task 1.3 report. See Task 2.1 baseline for ensemble analysis and Stage 2 recommendations.

4.4 KiB Raw Blame History