GEMMA DEVELOPER AGENTSAVED EVIDENCE
INTERNAL RUNS / SAVED EVIDENCE

Internal Runs

Compare saved attempts, then inspect the evidence behind each outcome. Local evaluation results are not Kaggle scores.

3 saved runs · Open a run to inspect individual attempts
Run / sourceModel / adapterEvaluation setSolved / recordedSource time
31B real rollouts · 480 (exact stack, 32 lanes, 13-Nov)local rollout bench (not a Kaggle score)Gemma 4 31B it, QAT W4A16 (gemma-4-31b-it-qat-w4a16-ct)rank-16 LoRA loaded with all-zero B weights, so it adds nothing (base model)60 SWE tasks x 8 rollouts (rollout31b-real test kit)261 / 480480 of 480 rollouts readable turn by turn2026-11-13 22:19 to 2026-11-14 00:07 New York time (the complete second attempt)
Archived E4B comparison · anchor_gpu0local evaluationUnknownAdapter unknownUnknown5 / 152026-11-12T04:40:11.993789+00:00
Archived E4B comparison · candidate_gpu1local evaluationUnknownAdapter unknownUnknown6 / 152026-11-12T04:40:11.993860+00:00