This baseline is a documented reconstruction of Yu et al., Designing Instance-Level Sampling Schedules via REINFORCE with James--Stein Shrinkage, version 2. It is not author-provided code, a GICO policy, or an exact reproduction of the paper's backbone and reward experiments.
The policy follows Table 4: image-shaped initial noise, native text tokens, optional pooled text, four convolution/cross-attention blocks, a feature pyramid, and a two-layer Dirichlet head with softplus plus 0.001. GroupNorm uses eight groups; cross-attention uses zero dropout. These unspecified numerical choices are explicit implementation decisions. Native input dimensions follow the backbone; report the actual parameter count rather than assuming 20 million.
Training uses a joint schedule action, two rollouts for each prompt/noise context, and the empirical baseline in Eqs. (4)--(10). In Eq. (10), the context-level cross-context quantity is implemented as the mean of that context's per-rollout leave-one-out baselines. The zero-variance case has shrinkage zero. Rewards and baselines are detached; log likelihood is taken over the complete Dirichlet action. There is no PPO clipping, KL penalty, entropy term, or reward-based selection at inference. The paper's AdamW settings are weight decay 0.0001 and gradient norm clipping at 1; learning rate and finite trajectory budget must be recorded by each experiment.
The paper gives L+1 Dirichlet intervals, L interior times, and a skipped interval, but also calls the last interval a stopping margin. Without the authors' sampler code these descriptions do not uniquely specify execution. This implementation uses the L interior times as model-evaluation times and appends terminal zero. It skips the interval before the first evaluation, leaves the sampled initial Gaussian unchanged, and executes exactly L Euler updates to zero. It does not add an uncounted evaluation at the original initial endpoint.
For a raw interval vector q of length L+1, solver intervals are
(1 - (L+1)*delta) * q + delta, with delta = 2**-20. This declared numerical
guard preserves a strictly ordered float32 grid. Raw actions are retained and
their original Dirichlet likelihood is used for REINFORCE; training and inference
use the same deterministic realization. No draws are discarded or retried.
The published empirical shrinkage coefficient uses the current batch rewards. Although the component means exclude each scored rollout, detaching the resulting baseline does not make its coefficient independent of that rollout. Accordingly, this reconstruction does not claim that the empirical gradient is unbiased or that the paper's theoretical MSE statements establish improved policy performance. It reproduces the stated empirical equations rather than silently substituting a different cross-fitted estimator.
Compare frozen backbone weights, guidance, resolution, actual NFE, terminal utility, evaluation prompts and generation seeds. Count every training rollout and reward call, plus separately disclosed calibration and validation costs. Preserve noise conditioning and the original policy family for this comparator; report its extra information and capacity relative to GICO. Such a comparison tests the complete methods, not the optimizer in isolation. Report preliminary validation results separately from locked-test benchmark results.
Run python -m genode.latent_clock.js_experiment fit --config experiment.json,
then use validation with exactly the same configuration file. Both commands
use separate generator and scorer devices. Each output phase must be new;
interrupted runs retain their observation journal and must use a new output
directory. The runner does not automatically resume partially completed fitting.
Configuration requires paths for runtime_config, manifest, anchor_evidence
(prepared shared GICO rows), anchor_measurements (original scored rows),
anchor_scoring (their scoring metadata), reward_calibration (the frozen
shared RewardCalibration payload), scorer_python, and output. It also
records nfe, two noise_seeds, and fit: trajectories, anchor_trajectories,
seed, batch_contexts, rollouts, and learning_rate.
Anchor reuse requires audited_implementations, an ordered pair of exact source
hashes for the measured anchors and this runtime. Audit the numerical generator
path before allowing a source pair; matching weights alone is insufficient.
Prompt/reference identities, native embeddings, solver settings, exact uniform
grids, and live scorer fingerprints/versions must also match. Both source hashes
remain in the artifact. Validation requires the same configuration, input file
hashes and runtime binding as fitting.
One uniform anchor is charged for each training prompt/noise context. Remaining trajectories must fill complete B-by-K batches. Rewards use the same frozen paired ImageReward/VQAScore calibration as GICO, including its scalar reward scale. For example, 128 anchors plus 48 batches of 32 contexts and two rollouts equals 3,200 trajectories. Calibration costs are additional and reported separately. There is no running reward normalization or utility surrogate-based JS objective.
The genode_js_reinforce_v1 artifact stores native architecture dimensions,
solver interpretation, numerical interval guard, immutable experiment identity,
training splits/seeds, reward/scorer provenance, runtime fingerprints and training
history. Loading verifies a manifest digest and uses a weights-only reader.
Ordinary inference loads only the JS policy, supplies native inputs, draws once
with sample_intervals, and calls materialize_js; rewards are never consulted.