INTACT-no-previous-action/README.md

INTACT without previous-action input

This directory contains 12 independently trained, task-specific E1 checkpoints whose intent actor is history-free in both training and evaluation.

Actor and evaluation contract

• Actor input: [z_t, m_t, z_t * m_t]
• No previous-action channel, initializer, expert prefix, or dataset-mean token
• 7/7 supervision window, 8 frames, terminal goal z7
• Full data, one epoch, batch 256, BF16 mixed precision, Math SDPA
• Official Direct evaluation, H=5, RH=5, zero search

Official Direct results

Task	Mean SR (%)	Train-seed sample std (pp)
PushT	85.33	0.58
Cube	99.44	0.38
Reacher	93.44	0.38
TwoRoom	98.78	0.69
Macro	94.25	0.08

The matrix contains 36 evaluations: training seeds 0/42/3072, evaluation seeds 0/1/42, four tasks, and 100 episodes per cell.

Files

• intact-goal-e1-single-history-free-seed<seed>.tar.gz: four task shards
• MANIFEST.json: release contract, aggregate results, and shard hashes
• RESULTS.csv: complete 36-cell history-free evaluation
• SHA256SUMS: archive checksums