# Independent grading audit

Original rubric re-audit of completed saved Qwen4 and Sol output stages; no network/model calls or candidate-code execution by this reviewer. Parent performed separately reported local T03 artifact replay.

Original rule: **>=16/20, every criterion >=3, no critical failure; minor edits count as accepted.**

| Task | Qwen initial → revision | Sol initial → revision |
|---|---|---|
| T01 | 18/20 (accepted) → 18/20 (accepted) | 20/20 (accepted) → 20/20 (accepted) |
| T02 | 19/20 (accepted) → 19/20 (accepted) | 20/20 (accepted) → 20/20 (accepted) |
| T03 | 19/20 (accepted) → 19/20 (accepted) | 17/20 (major repair) → 20/20 (accepted) |
| T07 | 14/20 (major repair) → 15/20 (major repair) | 20/20 (accepted) → 20/20 (accepted) |
| T08 | 19/20 (accepted) → 19/20 (accepted) | 20/20 (accepted) → 20/20 (accepted) |

**Matched first-stage acceptance: 4/5 each. Revised-stage acceptance: Qwen 4/5, Sol 5/5.** Qwen T09 initial is unpaired:19/20, accepted with minor edits. No critical-failure trigger found.

## Material findings

- **T01-R1-qwen4:** Summary labels other exclusions final despite unresolved orphan RF04; RF05 is correctly blank/provisional in context. Initial summary overgeneralizes which product is affected; revision clarifies Workshop unaffected.
- **T02-R1-qwen4:** Both stages call O5 stronger-GPU without supporting performance evidence. Correct recommendation/costs/eligibility otherwise.
- **T03-R1-qwen4:** Handoff retains unsupported original CAD NaN description; initial session-count discrepancy. Code functionality independently replayed by parent, not this reviewer.
- **T03-R1-sol:** Initial turn-limit output lacks required handoff; revision completes it. Functional code passes parent replay.
- **T07-R1-qwen4:** T704 purchase date October10/day10 is wrong: source October9/day11; exclusion at exactly20% remains correct. T708/T710 refund_cad=0 conflates no decided refund with unknown amount; notes explain convention but machine-readable triage remains materially ambiguous. T708 says Order/contact confirmed privately although no confirmation evidence exists; revision leaves this false action intact. Initial T705 invents end of course access and approval event; revision removes them.
- **T08-R1-qwen4:** Both stages claim assistant availability is per supplied CSVs although no assistant calendar is supplied; agenda remains feasible subject to that assumption.
- **T09-R1-qwen4:** Invoice1 shipping CAD20 is absent from PO. Note acknowledges verification gap but calls it nonblocking and says no discrepancies/OK to release; should obtain approval rather than infer authorization. Invoice footer establishes shipping tax treatment, not that the charge was expected or authorized.

## Inventory and limits

21 completed output stages audited: 20 paired stages on T01/T02/T03/T07/T08 plus Qwen T09 initial. Qwen T04 and Sol T09 have partial transcript material, not completed deliverables. T05/T06/T10 have no completed stages for either model. Every second repetition is absent. Qwen T09 revision absent. Full 80 planned stage inventory is in JSON.

T03 functional criteria incorporate the parent’s reported 19/19 local replay on all four saved sites; this worker only read artifacts and handoffs. Sol initial still fails the handoff criterion despite functional code. No candidate code, model APIs or network used by this worker.

Read both synthetic invoice images directly and recomputed invoice math. Parsed actual CSVs, checked source policy and output text. Original task8 UTC conversion is +7 hours; its rubric wording must not be interpreted as output timezone UTC+7.

Existing independent-review files lag saved Qwen T03/T08 revisions and T09 initial; those are now graded. Under illustrative all-criteria-4 sensitivity, paired Qwen is0/5 at both stages and Sol4/5 then5/5. This is stricter than original acceptance and must be labeled separately.

This is unblinded AI grading of five one-run matched synthetic tasks, not human or blinded testing, reliability estimation, or measured human-editing time. Forced revisions cannot establish adaptive cost-to-success.
