JARVISLabs

Local model on the frozen Gate C battery (bounded adapters)

Passing exact cases is not enough: every run that improved exact lost ground on transfer or anchors.

Local model on the frozen Gate C battery (bounded adapters)

Qwen3.5-9B, 48 updates 11 Gemma base (no adapter) 4 Gemma exposure 48 9 Gemma data diversity 10 Gemma plan010 (coverage) 8
Scale: passes, out of 12. Gate C: ≥10 exact passes and 0 fails

Twelve bounded experiments in two days. Each froze its plan, reloaded the adapter in a fresh process and was graded on the same 60 frozen cases. None passed Gate C.

Test: Exact partition of the frozen 60-case battery (12 exact, 40 transfer, 8 anchors); greedy decoding; independent agent reviewers.

Protocol

Test definition

Exact partition of the frozen 60-case battery (12 exact, 40 transfer, 8 anchors); greedy decoding; independent agent reviewers.

Limits

Limitations

  • Graders are independent agent reviewers, not blinded human qualification.
  • Only the exact partition is charted; transfer and anchor results are in each point’s note.

Log

Recorded values

Variant Value Evidence Samples Source
Qwen3.5-9B, 48 updates 21 Sept 2026 11 passes TESTED 12 —
Gemma base (no adapter) 22 Sept 2026 4 passes TESTED 12 —
Gemma exposure 48 22 Sept 2026 9 passes TESTED 12 —
Gemma data diversity 22 Sept 2026 10 passes TESTED 12 —
Gemma plan010 (coverage) 23 Sept 2026 8 passes TESTED 12 —

Source paths refer to the private JARVIS repository and runtime records. They are listed for audit and are not published.