Five-pass latent-video temporal refinement

You inherit a deterministic policy that leaves a noisy short-sequence latent trajectory unchanged. Improve its temporal refinement schedule while preserving genuine motion and shot boundaries. A trusted evaluator applies the policy to disjoint procedural panels and compares the refined sequence with a clean reference trajectory.

Hard Constraints

• Edit only /app/methods/main/policy.py; use /app/experiment_log.md to record experiments.
• Keep build_policy() -> dict as the entry point.
• Return exactly the temporal-denoise-policy-v1 mapping documented in /app/API.md.
• The five pass logits must be finite Python floats in [-4, 4].
• The remaining controls must be finite Python floats within the bounds in /app/API.md.
• Do not return trajectories, arrays, callbacks, paths, source code, or extra keys.
• Do not inspect or interfere with another container, Docker, credentials, verifier files, or process controls.
• The verifier has no network. The submission must be deterministic and use only the Python standard library.

Run the local contract and determinism checks with:

python /app/selfcheck.py

Run a complete visible evaluation with:

python /app/score_visible.py

What You Have

• Three public panels spanning six balanced families of short latent-video trajectories.
• /app/API.md, which specifies policy normalization, five-pass accounting, gate behavior, and the visible metric.
• A weak no-refinement policy in /app/methods/main/policy.py.
• Deterministic visible feedback. Public panels are development feedback only and are disjoint from verifier panels.

What You Submit

Harbor collects policy.py and the experiment log. The verifier stages policy.py, executes build_policy once in a constrained no-network child, validates the small numeric mapping, terminates that child, and only then evaluates the frozen policy on sealed panels.

How It Is Judged

Lower balanced normalized temporal latent error is better. Each panel-family unit is scored independently and the units are averaged equally, so a schedule must transfer across static content, camera-like drift, acceleration, oscillation, transient corruption, and shot-like changes. The trusted runtime performs exactly five temporal refinement passes per trajectory.

Common Pitfalls

• Returning integers, booleans, numeric strings, NaN, infinity, or extra keys.
• Confusing pass logits with already-normalized pass fractions.
• Applying too much smoothing and erasing real motion or boundaries.
• Using gates so aggressively that transient corruption is preserved.
• Overfitting one visible family instead of checking all panel-family units.