BASE MODEL 01
—
exact-match accuracy
— false completion calls
01 / THE QUESTIONOPEN-WEIGHT DECISION LAB
Did the agent
actually finish?
A draft is not a send. A tool call is not proof of a result. Compare GLiNER2.5-Decide before and after local fine-tuning on the gap between attempted and done.
BASE VS TUNEDSANDBOX-SCORED CASESINSPECT EVERY ERROR
02 / THE EVIDENCE
LOADING REPORTThe held-out result
AFTER FINE-TUNING 02
—
exact-match accuracy
— false completion calls
MEASURED CHANGE Δ
—
percentage points
— changed decisions
Loading study notes…
03 / THE CASES
— CASESInspect every decision
Loading the recorded run…
04 / REPRODUCE ITOne model.
One model.
A measurable definition of done.
The GLiNER2.5-Decide checkpoint runs locally. A rank-8 LoRA adapter trains on versioned examples; the best checkpoint is selected on development loss and compared with the base on task-template-held-out cases.
uv run --extra model exitreceipt train --data data/v2-cases.psv --evidence data/v2-evidence.psv --run-dir runs/v2 --device mps --epochs 3 --lora-dropout 0.1Model revision, data hash, software versions, settings, and every prediction are recorded. See the v2 experiment contract for the exact method and limits.