# H3 causal-ICL v1 — paired data and completed diagnostic

**配对视频审核入口（2026-10-18）：** [每对视频相邻存放、两列动图与 MP4、分页浏览及 12 组抽查备注](pair_review/README.md)。原配对、媒体和标签保持不变。

**2026-10-18 first-frame correction:** the historical 12-video run used first-frame attention conditioning, without a per-step clamp on the target first temporal latent. It is not strict temporal anchoring. Original inputs/results are unchanged; see [the dated erratum](audit/first_frame_control_erratum_20261018.json).

This release adds the current causal-rule pairbank and the completed outcome-hidden small diagnostic. It is independent of the older `physics_cards_v7` retrieval export. [Browse the code](code/physical_icl_h3/README.md) or download [the code ZIP](code/physical_icl_h3_code.zip). The matching GitHub repository is [Physical-VICL](https://github.com/chengye629/Physical-VICL).

| Artifact | Contents |
|---|---|
| `pairing/` | 5,341 directed assessed pairs; 2,852 rulebank entries; video splits; original judgments and metadata-only response screening |
| `training_inputs/` | Exact 65 train / 2 dev task specifications, query-only and matched-demo prompts; no response rubrics |
| `quality/` | Original selected pair assessments, 57 response-compatible train candidates, 8 mechanism-only review cases, sample storyboards |
| `results/` | Both actual configs, 200-step losses, actual query order, initialization and final checkpoint receipts |
| `videos/` | 12 final-checkpoint videos and actual-input sidecars |
| `evaluation_only/` | Dev GT, positive and irrelevant clips, response/action rubrics and unfilled human ratings |
| `previews/index.html` | Six results per query plus evidence clips; download the folder to use the local HTML review page |
| `audit/` | Independent engineering acceptance and preliminary assistant visual observations |
| `benchmark_design/` | WISA100 training-support candidates and Base review template; no invented Base-failure labels |
| `provenance/` | Original source ZIP, actual tokenizer strings, environment, source hashes and experiment assumptions |

Both independently initialized models trained for 200 updates on the same 65 queries in the same order, with the same target-noise schedule. Learning rate 1e-4, warmup 10, linear decay; LoRA rank/alpha 16; effective batch 1; seed 42; 576x320x124/24 fps. Two models x two dev queries x three conditions = 12 videos. No new judge model was deployed.

**Scope:** pipeline/conditioning diagnosis, not a population benchmark. There are only two dev queries and no unadapted H3 Base outputs in this 12-cell run. Query-only SFT is a trained model, not Base. Training losses cannot establish causal physical-rule transfer.

**Quality:** `positive` in the historical pairbank is a broad mechanism label, not guaranteed same-response evidence. Eight of the 65 training pairs have different responses in the original model judgments. They were used in the frozen run and remain recorded. The separate 57-row list is a conservative metadata screen, not a human-verified replacement dataset. Original false/unverified flags are preserved. Near-duplicate content checks remain pending.

**Crops:** the run explicitly assumes the target event is in the existing first 124 frames at 24 fps. This assumption has not been upgraded to visual verification. Some sample storyboards warrant timing/response review. No historical input, crop or checkpoint was changed during acceptance.

**Paths:** JSON exports use project-relative paths. Checkpoint references for the later outcome-hidden models identify the original experiment; their weights are not duplicated here. The separate first ordinary SFT final adapter is hosted under `ordinary_pilot_v1/checkpoint-0000200/`. Original train source videos and all query frames are not duplicated; the six dev evidence videos are included for inspection. Resolve WISA source paths against a local video root. `provenance/outcome_hidden_tasks_v1.original.zip` is the byte-exact historical package with its original paths; it is provenance, not the default training loader.

**Model input vs evaluation:** `training_inputs/` contains outcome-hidden task prompts. Rubrics and full observations live separately in `evaluation_only/`. Exact encoder strings include native media headers/time labels plus the approved h3_prompt, never rubrics or original full captions.

**Reuse terms:** existing upstream WISA/media terms continue to apply. This release does not grant new rights over third-party videos. Trainer code has its own Apache-2.0 license in the code mirror.

## Complete task bank and first SFT — added 2026-10-17

- [All 345 task records](outcome_hidden_tasks_v1_all345/README.md): 320 train / 25 dev, all drafts, 345 first-frame previews, all primary/continuation candidates, separate evaluation evidence and a 345-row quality audit. The 46 held tasks are retained with reasons; they are not silently enabled for training.
- [First ordinary SFT](ordinary_pilot_v1/README.md): original 320-row inputs, actual 200-step logs/losses, final LoRA adapter, 32 historical Base/LoRA videos and the separate q04 controlled probe.


## Training-focused follow-up — 2026-10-18

[Training code, scale ledger, validation panel and CPU status](research_v2/README.md) preserve the existing assessed pairs and prepare strict-first-frame / training-objective controls. Expanded default: 277 existing outcome-hidden train queries, 3 complete passes (831 updates/arm). No new VLM review or GPU result is claimed. The future pairing-selector proposal is not part of this training job.