File size: 2,925 Bytes
b8da34b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 | # PAWBench A-09 ratio-control candidate-video staging (Issue #298)
Collaborator-writable staging surface for the A-09 model-level ratio-control
case study ([PhysEdit#293](https://github.com/Andrew0613/PhysEdit/issues/293),
execution issue [#298](https://github.com/Andrew0613/PhysEdit/issues/298)).
This repo realizes the `candidate_data` location proposed by
`staging_locations.json` in the #297 input package; it lives under the
`PhysEdit` org (public) by operator decision of 2026-07-13 because the proposed
`Andrew613` repos were never created and private quota was unavailable.
## What this repository contains
- `candidate_videos/` — Seedance 2.0 first/last-frame candidate training
videos for scene A-09, named by frozen `sample_id`. The formal denominator
is exactly 200 rows (40 endpoint controls x 5 seed identities). Videos
appear here as soon as they are generated, **before** any usability
filtering, per the #298 upload-before-exclusion contract.
- `manifests/candidate_ledger.jsonl` — one terminal-status row per planned
sample (schema: `candidate_ledger.schema.json` from the #297 package),
updated at each phase boundary.
- `receipts/` — phase receipts, launch authorization (with live price quote,
provider schema deviations, and the fal-storage input-transport fallback),
and the hash-bound fal staging map.
- `review/` — canary visual-review contact sheet and motion strips.
## What this repository is NOT
- These are **candidate training-data videos**, not PAWBench benchmark
inputs, not accepted training data (acceptance is a separate reviewed
gate), not model outputs under evaluation, and not evidence that any
training approach achieves probabilistic alignment.
- No usability decision is implied by a video's presence here. The accepted
bank and failure manifests are separate, later artifacts.
- No PAWEval result, checkpoint, metric, or paper-facing claim lives here.
## Provenance
- Input authority: `Andrew613/PAWBench-A09-LastFrames` @
`800521f1cc6332cd834632a1ecdecb81b17937e4` (normalized A-09 first frame,
40 terminal-frame controls, frozen 200-row generation plan, frozen
Seedance 2.0 request profile).
- Generator: fal.ai `bytedance/seedance-2.0/image-to-video`, standard tier,
720p, 5 s, `aspect_ratio=auto`, audio disabled, direction-neutral frozen
prompt. The provider exposes no seed parameter; planned seed identities
are sample identities (provider-returned seeds are recorded in the run
sidecar).
- Runner: `projects/pawbench/tools/a09_ratio_control/run_seedance_candidate_grid.py`
on branch `task/pawbench/0298_seedance_bank` of PhysEdit.
- Every video is hash-bound to its plan row (first-frame hash, endpoint
hash, prompt hash, request-profile hash) in the candidate ledger;
`SHA256SUMS` covers the uploaded files.
No credentials or secrets are stored here; receipts record credential
environment variable names only.
|