djduecaster/dfreq-sprint-reliability-20261107

dfreq Sprint reliability research policy

A genuine stateless feedback policy selected through behavior-diversity research. Prepared for review as a distinct daily-points entry. Nothing is published or submitted by this package.

Changes and provenance

Starting from the previously trained robust Sprint network, this candidate suppresses the residual branch input weights and changes the main network's first-layer weights and bias. The main network and remaining feedback remain observation-dependent. The combined change has independently confirmed reliability gains; it is not attributed to branch suppression alone. This is not a renamed copy, cached-action policy, seed recognizer, or replay lookup.

The original neural policy is djduecaster/dfreq at revision 7b67892a603eb95afb0ffbfe61d07b7d57128f12. Its upstream lineage and contributor credits are retained in upstream-base-card.md. The original teacher lineage included clock/seed mechanisms, but those mechanisms are absent from this deployed graph. No third-party model was downloaded or executed for this preparation.

Local validation

The unchanged official Arena evaluator selected command 1.10 m/s and returned 0.959 seconds, 5/5 finishes: 0.950, 0.985, 0.960, 0.965, 0.935 seconds. Two historical sweeps and a fresh November 7 Linux sweep agree exactly. This is local simulation, not an official hosted score.

Independent reliability tests at fixed command 1.15 finished 689/1,024 versus 525/1,024 for the prior robust model, then 659/1,024 versus 549/1,024 in separate confirmation. This is a reliability gain, not a speed record. Broad-start reliability remains imperfect.

The graph passes ONNX validation, uses float32 obs[1,61] to actions[1,14], has no external tensors or custom operators, and produced deterministic finite actions on 2,538 real observations. No clock, seed ID, hidden actor state, lookup, or privileged deployed observation is used.

Scope and limits

This is a daily-publication-points candidate, not a personal-best replacement. It is slower than the existing official 0.848-second entry, and its prior speed gate remains failed. The original research-retention status is preserved in the evidence folder. Preparing this package does not authorize publication or submission.

Simulation only. No physical-robot tests or safety guarantee. ONNX output amplitudes are not a physical deployment validation. Arena acceptance, hosted score, and points are not guaranteed. A non-Tug submission does not change the selected Tug Cup policy.

The inherited repositories do not declare a standalone model license. No MIT, Apache, or other blanket license is asserted. Retain upstream credits and review the applicable upstream and Hugging Face public-repository terms before reuse.