polyphony · gemini-2.5-flash

✗ fail↓ regression 100% → 86%

Pass rate

86%

Cases

7 (1 failed)

Mode

eval

Suite

dracula

Env

Started

2026-07-13 09:52 UTC

Duration

—

Cost

—

Tokens in/out

— / —

Commit

—

Branch

—

Source

—

Labels

legacy_env	ci-dracula

Compared with the previous run

Baseline 2026-07-13 05:21 UTC · 100% → this run 86%

cases added / removed

+ eval:attribution
+ eval:continuity
+ eval:outline
+ eval:prose

Cases (7, 1 failed, failures first)

eval:attributionscore 0.33✗ fail

Output

accuracy=0.3333 chance=0.3333 n=3

eval:continuityscore 0.50✓ pass

Output

detection_recall=0.5 fpr=1.0

eval:extractionscore 0.71✓ pass

Output

F1=0.7059 P=0.6 R=0.8571 predicted=['Aldous Kerr', 'Nora', 'Percy', 'Mr. Boone', 'Cora', 'Dr. Ward', 'Sister Agatha', 'Renwick', 'Verhoeven', 'Elias']

eval:ingestionscore 0.90✓ pass

Output

grounded=9/10 avg_chunks=21.5

eval:outlinescore 0.60✓ pass

Output

nodes=6 structural_ok=True

⚖︎ Judge rationale

The outline recovers some of the reference beats, particularly the inciting incident and climax, but significantly alters the order and nature of the rising action and midpoint, making it a different story shape.

eval:prosescore 0.60✓ pass

Output

words=701

⚖︎ Judge rationale

The scene fails to dramatize the requested beat of the characters arguing about pressing on, instead focusing on a philosophical discussion about truth and obfuscation. While the prose is coherent and readable, and the characters have distinct voices, the core dramatic requirement is not met.

eval:retrievalscore 0.82✓ pass

Output

precision@k=0.8182 mrr=0.9545