# Source inventory and authority

## Authority order

1. Server run artifacts and full configs on `jlu66`.
2. This corrected handoff package.
3. Public development repositories: [ReTime_source](https://github.com/miaochengy-pngu/ReTime_source) and [OPTR-StarVLA](https://github.com/miaochengy-pngu/OPTR-StarVLA).

Server facts override stale prose or inferred settings in a repository snapshot.

## Server-authoritative origins

- Development tree: `/home/miaochengyu/starVLA-infoflow-codex`
- Curated source tree: `/home/miaochengyu/OPTR-StarVLA-clean`
- 20K parent: `/mnt/miaochengyu/starvla-infoflow-runtime/runs/florence-scratch-official-100k/florence-scratch-uniform-seed-42-official100k-mb4-eval1k-20260726`
- Confirmatory arms: `/mnt/miaochengyu/starvla-infoflow-runtime/runs/infoflow-confirmatory-22k-20260730`
- Frozen proposal: `/mnt/miaochengyu/starvla-infoflow-runtime/runs/infoflow-proposals/scratch20k_residual_grid8_gate42f_batch1_20260729/residual_temporal_proposal.json`
- Formal aggregate: `/home/miaochengyu/starVLA-infoflow-codex/logs/bridgert/infoflow_confirmatory_eval/seed_117465/final_test_policy_seeds_20p5/aggregate_summary.json`

## Corrections applied during handoff

- Corrected DeepSpeed from a stale ZeRO-3 description to the actually used **ZeRO-2** configuration.
- Corrected `diffusion_model_cfg.cross_attention_dim` from stale 2048 to the actually used **1024**.
- Fixed world size directly at 2 in the portable Accelerate config.
- Restricted the public launcher to the uploaded Uniform and Residual arms.
- Removed engineering checksum/path gates from launch.
- Kept the real semantic identity requirements for data, episodes, valid anchors and proposal support.

## Key implementation files

- `starVLA/training/train_starvla.py`: trainer, exact-state resume, checkpointing.
- `starVLA/training/infoflow.py`: InfoFlow/importance-correction integration.
- `starVLA/dataloader/temporal_proposal.py`: proposal parsing and support semantics.
- `starVLA/dataloader/infoflow_sampler.py`: traceable sampling.
- `starVLA/scoring/flow_residual.py`: residual scoring.
- `starVLA/scoring/residual_cache.py`: score-cache handling.
- `scripts/run_florence_bridge_arm.sh`: portable two-GPU continuation launcher.
- `scripts/evaluate_bridge_384_no_video.sh`: locked 384-rollout orchestrator.

The public dataset and base Florence model are not redistributed in this package. Their local paths are supplied through environment variables.