polyphony · claude-opus-5-5
✓ pass≠ not comparableMetrics
| prose.score | 0.630 (was 0.992) |
| prose.words | 3,659 (was 3,607) |
Labels
| cadence | on-demand |
| experiment | w17-r2-1 |
| harness_sha | 86123fdae7a0f3d2805e1ab75905d51c08993484 |
| judge_provider | groq+anthropic |
| judge_self_graded | True |
| repeat | 1 |
| trigger | workflow_dispatch |
Compared with the previous run
Baseline 2026-10-25 00:24 UTC · 100% → this run 100% · code diff ↗
- eval:prose: 0.99 → 0.63
Cases (1)
words=3659 path=planned rubrics={'R1': 0.8325, 'R2': 0.2375, 'R3': 0.9575, 'R4': 0.94, 'R5': 0.89} seams=0/0scene 1 R1: Nora Vance, listed in the planned beat cast, appears only as a portrait mention, so a principal is reduced to a mention, meeting the 0.4 level. | scene 1 R2: With only three untagged lines, the evaluator cannot attribute ≥50% correctly, making voices indistinguishable. | scene 1 R3: All planned beats appear sequentially, each change occurs as described, with no extra or missing events. | scene 1 R4: The passage maintains consistent first‑person‑limited POV and past tense, matches the intended tone, and presents varied, concrete prose without any structural violations. | scene 1 R5: The prose contains no contradictions to the canon or prior text and follows the planned beats without error. | scene 2 R1: All principal canon characters appear with correct names and behavior, matching the planned cast exactly. | scene 2 R2: No dialogue was presented to evaluate voice distinctness, so the rubric cannot be applied. | scene 2 R3: All five planned beats appear sequentially with their required changes and no unplanned material. | scene 2 R4: The scene maintains consistent POV and tense, matches the requested tone, and is a polished narrative without structural errors. | scene 2 R5: The scene aligns fully with the canon and planned beats, containing no contradictions or continuity errors. | scene 3 R1: All principal characters (Aldous Kerr, Count Vasska, Nora Vance) appear with correct names and act according to canon, with no missing or mis‑cast figures. | scene 3 R2: Around half the untagged lines can be matched to the correct character, but voices often drift toward generic narration rather than distinct sample styles. | scene 3 R3: All four beats are realised in order with their stated changes landing (the reluctant agreement, the dictation and forward-dated letters, the failed hidden warning, the sealed letters carried off), with no later beat anticipated; minor additions such as the wax sealing and the map of his coast are s | scene 3 R4: The scene keeps close third-person past tense from Aldous's POV throughout, matches the Gothic register and prior text, and uses concrete, varied sentences with no labels or summary; it loses a little for a few minor slips (the Count carrying the letters 'toward the door' while still holding the can | scene 3 R5: The scene keeps to the canon and prior text: names, the locked-in situation, the dictated letters, the nocturnal Count and the island coast all fit, and it follows the planned beats. The only quibbles are small unsupported details, namely 'the grave young lady in the portrait' and the candle appeari | scene 4 R1: Only Aldous Kerr and Count Vasska appear, as the beat cast specifies; the name "the Count" is used correctly, and both act in line with the brief (Aldous resolves, panics and recognizes his captivity; Vasska is courteous and menacing), with no invented characters and only a small embellishment (the | scene 4 R2: Lines 1 and 2 are clearly Count Vasska ("my friend," "I would not have you," "You shall not"), but line 3, "A prisoner.", is too short and generic to show Kerr's diction or cadence, so it is attributable only by context or elimination, and the sample is too thin to show voices fully distinct. | scene 4 R3: All five planned beats appear in order with their stated changes landing (cheerful resolve and echoing steps, door-trying with unsteady hands, sheer-drop window realization, the Count's courteous warning, naming himself a prisoner and searching the room), with only minor unplanned details such as th | scene 4 R4: The scene keeps close third-person past tense from Aldous's POV, matches the sample's Victorian-gothic tone and rhythm, and follows all five beats as dramatized scene with no labels or prefixes, though a few phrases are generic or conventional ("as a man walks who has an errand," "breathes as a man | scene 4 R5: The scene follows the brief and prior text closely, with one minor slip: Aldous stands against the door at the foot of the stair, but la