H3 Omni Ref Iterations

Baseline 20

Original accepted-20 run, with the current evaluator applied to the 17 prompt-valid cases.

Current strict QA: 12 / 17 accepted

20 cases

Round 1 · Prompt A/B

Direct vs minimal six-section, matched seeds, with final soundtrack delivery where applicable.

2 accepted / 10 after final delivery

10 cases

Round 2 · Matched Seeds

Compact contract vs minimal six-section; deterministic silence/source soundtrack.

Core 2 / 8 · expansion gate failed

10 cases

Round 3 · Official Format A/B

Five baseline-success combinations with matched seeds: direct requests vs official six-section prompts. Gemini 3.7 prompt QA approved all five pairs.

Core 8 / 8 · pair coverage 4 / 4

10 cases

Promotion 20 · Matched A/B

Twenty semantically accepted real-workload combinations. Direct and Context-IR-style prompts use identical references, duration, seed, and optimized 4xGB200 runtime.

Generated 40 / 40 · QA 11/40 accepted · both accepted 2, direct only accepted 3, neither accepted 11, structured only accepted 4 · task gates: A1 0% (3 pairs, needs prompt iteration), A3 50% (4 pairs, needs prompt iteration), A4 33% (3 pairs, needs prompt iteration), B3 50% (4 pairs, capability canary), D2 33% (3 pairs, capability canary), D4 100% (3 pairs, needs prompt iteration)

40 cases

Repair 15 · Base-Compatible

Fifteen coherent real-workload combinations, three each for A1, A3, A4, B3, and D4. Direct prompts use only FastVideo-bound media labels; unsupported Subject labels are removed.

Generated 15 / 15 · QA 7/15 accepted

15 cases

Audio Repair 5 · No Invented Speech

Five A3/B3 retries with the same reference workloads but an explicit non-verbal contract: mouths at rest, no spoken dialogue, and ambient room tone only.

Generating 0 / 5

5 cases