Ten-evaluation generative integrator policy optimization

You inherit a valid but deliberately simple policy learner for a frozen unconditional image generator. Improve the bounded integration policy returned by that learner; a trusted evaluator refits it from scratch on sealed paired endpoint tensors and evaluates a sampler that makes exactly ten frozen-model calls.

Hard Constraints

• Edit only /app/methods/main/optimizer.py; use /app/experiment_log.md only to record experiments.
• Keep fit_policy(calibration_dir, budget) -> dict as the entry point.
• Return only the policy-v3 mapping documented in /app/API.md. Never return images, tensors, model objects, callbacks, paths, or executable code.
• The returned mapping must use the required solver="generalized" contract value and must describe exactly ten model evaluations.
• Consume paths, limits, and seeds from the supplied budget. Do not hard-code evaluator paths or assume undisclosed filenames, examples, or seed domains.
• Do not inspect or interfere with the evaluator service, another container, Docker, credentials, cgroups, request queues, or verifier files.
• The fitting process has no network. All required dependencies and frozen assets are already available through the documented runtime.

For a learner that returns quickly, run the contract check with:

python3.10 -I /opt/ara-task/visible_client.py --optimizer /app/methods/main/optimizer.py --contract-only

The contract check executes fit_policy completely. It does not generate the 10k scoring images or consume the eight-call budget, but a training-heavy learner can still make it expensive; use it deliberately rather than before every full evaluation.

Run a complete visible evaluation by removing --contract-only. The client creates one bounded request and waits for the evaluator's read-only response; do not manipulate the transport directories directly.

At most eight complete visible evaluations are accepted per task run. Contract-only checks do not consume this budget. Keep the best completed candidate before the limit is reached; further full requests fail closed.

What You Have

• /app/visible/calibration/endpoints.safetensors contains 1,400 aligned finite float32[3,32,32] pairs under the keys noise and images. Treat them as input and target endpoints; their construction procedure is intentionally not part of the optimization surface.
• /app/API.md specifies every budget field, public helper, policy key, numerical bound, and serialization limit.
• The fit sandbox exposes the same neutral time-uniform, zero-correction policy returned by the inherited starter through budget["public_initializers_dir"]. It may be modified or ignored.
• model_io.load_frozen_model, TrainableGeneralizedSolver, and the strict policy validator are available inside the fit process.
• Visible evaluations use a fixed public split and return a deterministic lower-is-better distribution distance. They are development feedback only.

What You Submit

Harbor collects optimizer.py and the experiment log. The verifier stages only optimizer.py, executes it in a fresh no-network sandbox, validates the JSON policy, proves the fitting process has terminated, and then starts trusted generation. Keep the submission self-contained and deterministic under the provided fit seed.

Temporary probes, deliberate exceptions, and partially edited learners are never valid final artifacts. Before finishing, restore the optimizer from the best completed full visible evaluation and run one final contract check; a failed check does not automatically restore a prior candidate.

How It Is Judged

Lower distribution distance is better. The trusted sampler uses exactly ten frozen-model evaluations. Visible and sealed evaluations use disjoint examples and seed sets, and the sealed protocol uses a larger sample set; compare only measurements produced by the same protocol.