fwerkor/CID-RQ1-StageA-v0

CID RQ1 StageA-v0 artifacts

This artifact records the mechanism-feasibility validation of the original CID StageA-v0 checkpoint used for RQ1. It is intentionally not the final CID model. The backbone is GSAI-ML/iLLaDA-8B-Base, pinned in the experiment to revision a1b5b5f8a31a3854a46205ee584178c04b45ec9a.

The checkpoint learned useful downstream CID interaction heads, but the original end-to-end runtime failed to create information-need cells because randomized physical-slot supervision was paired with exact per-slot BCE and an absolute allocation threshold. The artifact therefore reports both the unmodified runtime and a validation-only allocation intervention that changes no model weights.

Key NQ-Open held-out results (32 examples):

• Original runtime: 0/32 tool executions, 0/32 fresh observations, 0% observation coverage.
• Allocation intervention at the original 128-slot geometry, no weight changes: 29/32 tool executions, fresh observations, and cognitive projections; 90.625% mean observation coverage; 24/32 converged.
• The same allocation intervention with training-aligned 8-slot geometry: 30/32 tool executions, fresh observations, and cognitive projections; 93.75% mean observation coverage; 28/32 converged.
• Teacher-forced head probe: source argmax 165/165; argument presence mean 0.99994; argument grounding cosine mean 0.8166; allocation positives mean 0.1098 and 0/219 exceed the original 0.8 runtime threshold.

The held-out NQ examples come from the pinned official nq_open validation parquet and are separate from the NQ-Open train split used by CID-Dataset. See data/registry.json and the raw JSONL for exact provenance. The artifact corrects a generator metadata label from train to validation wherever the pinned upstream row key already points to the official validation parquet; this metadata-only correction does not change any model input or result.

Known StageA-v0 issues are documented in summary.json. RQ2 and later experiments use a corrected train/runtime contract and a model retrained from the base checkpoint.