Scientific purpose
The central Prob4D external-validity gap is still real provider competence: the controlled probabilistic mechanism is positive, but no frozen real dynamic-geometry provider has yet produced a successful claim-bearing source path. The frozen CUT3R version is already terminal for its registered attempt and must not be silently retried or reinterpreted.
Register a new provider version based on the official MonST3R dynamic-geometry implementation. This is an independent experiment, not a rescue of CUT3R.
Frozen upstream identity
Initial upstream to freeze before any source run:
- repository:
Junyi42/monst3r
- upstream revision:
574cc77ad278bad582f470e5382624e01f8769a7
- provider family: MonST3R / dynamic point maps and camera estimates for dynamic video
- checkpoint identity: must be content-hashed on the trusted runner before execution; the workflow must fail closed if the expected hash is absent or mismatched
- license/provenance: record the exact upstream model/code license in the run artifact
If the upstream revision/checkpoint identity changes, that is MonST3R provider v2, not a retry of v1.
Information-order rule
This experiment is source-only. It may use only the already opened provider-development/source cohort. It must not read, enumerate, score, or infer from the protected held-out target cohort. No thresholds, frame selection, postprocessing choice, uncertainty scale, or candidate policy may be selected from target outcomes.
Trusted-runner rule
Do not download and execute arbitrary external code from an untrusted PR job. Follow the existing trusted self-hosted pattern:
- protected/pre-provisioned local MonST3R checkout at the frozen revision;
- protected/pre-provisioned checkpoint with verified SHA-256;
- authorized exact Prob4D revision/head;
- NVIDIA preflight and recorded GPU/driver/CUDA environment;
- source cohort mounted read-only;
- final artifacts uploaded even on provider failure.
The workflow should refuse to run when any protected dependency, exact revision, model hash, or source mount is missing.
Stage 1: source-only provider feasibility
For a small frozen source clip roster, record per clip:
- input frame IDs/hashes and exact preprocessing;
- number of frames requested/decoded/forwarded;
- whether provider inference entered and completed;
- finite dynamic 3D outputs / valid masks / camera outputs;
- output coordinate convention and scale handling;
- runtime, peak GPU memory if available, model/upstream/checkpoint IDs;
- exact exception/failure stage if unsuccessful.
Success gate: at least one preregistered source clip must complete inference and emit finite, nondegenerate dynamic geometry that can be deterministically converted into the Prob4D observation contract. Merely loading weights or entering a CUDA forward call is not success.
If the gate fails, retain the complete failure artifact and mark MonST3R provider v1 terminal. Do not mutate the frozen run until it passes.
Stage 2: adapter qualification (only after Stage 1 success)
Open a separate PR that freezes the conversion from MonST3R outputs to the Prob4D input contract. Qualify it on source data only with deterministic replay, dimensional/coordinate assertions, finite-value checks, and source-side predictive diagnostics. Do not open the target cohort in that PR.
Paper claim boundary
A successful source-only MonST3R run would establish that Prob4D is not limited to controlled synthetic uncertainty and that a second real dynamic-geometry family can physically produce admissible source observations. It would not by itself establish target transfer, calibration, downstream BayesianPhysTwin benefit, unseen-object generalization, or state of the art. Those require separately frozen target evaluation.
This experiment is secondary to the BayesianPhysTwin/Causal4D cross-action transport result: it strengthens external validity, whereas cross-action transport tests the paper's core physicality claim.
Scientific purpose
The central Prob4D external-validity gap is still real provider competence: the controlled probabilistic mechanism is positive, but no frozen real dynamic-geometry provider has yet produced a successful claim-bearing source path. The frozen CUT3R version is already terminal for its registered attempt and must not be silently retried or reinterpreted.
Register a new provider version based on the official MonST3R dynamic-geometry implementation. This is an independent experiment, not a rescue of CUT3R.
Frozen upstream identity
Initial upstream to freeze before any source run:
Junyi42/monst3r574cc77ad278bad582f470e5382624e01f8769a7If the upstream revision/checkpoint identity changes, that is MonST3R provider v2, not a retry of v1.
Information-order rule
This experiment is source-only. It may use only the already opened provider-development/source cohort. It must not read, enumerate, score, or infer from the protected held-out target cohort. No thresholds, frame selection, postprocessing choice, uncertainty scale, or candidate policy may be selected from target outcomes.
Trusted-runner rule
Do not download and execute arbitrary external code from an untrusted PR job. Follow the existing trusted self-hosted pattern:
The workflow should refuse to run when any protected dependency, exact revision, model hash, or source mount is missing.
Stage 1: source-only provider feasibility
For a small frozen source clip roster, record per clip:
Success gate: at least one preregistered source clip must complete inference and emit finite, nondegenerate dynamic geometry that can be deterministically converted into the Prob4D observation contract. Merely loading weights or entering a CUDA forward call is not success.
If the gate fails, retain the complete failure artifact and mark MonST3R provider v1 terminal. Do not mutate the frozen run until it passes.
Stage 2: adapter qualification (only after Stage 1 success)
Open a separate PR that freezes the conversion from MonST3R outputs to the Prob4D input contract. Qualify it on source data only with deterministic replay, dimensional/coordinate assertions, finite-value checks, and source-side predictive diagnostics. Do not open the target cohort in that PR.
Paper claim boundary
A successful source-only MonST3R run would establish that Prob4D is not limited to controlled synthetic uncertainty and that a second real dynamic-geometry family can physically produce admissible source observations. It would not by itself establish target transfer, calibration, downstream BayesianPhysTwin benefit, unseen-object generalization, or state of the art. Those require separately frozen target evaluation.
This experiment is secondary to the BayesianPhysTwin/Causal4D cross-action transport result: it strengthens external validity, whereas cross-action transport tests the paper's core physicality claim.