Skip to content

Measure native grip interruptions and document goal-image control - #22

Open
Tarzelf wants to merge 3 commits into
research/retention-baseline-20261016from
feat/native-contact-audit
Open

Tarzelf wants to merge 3 commits into
research/retention-baseline-20261016from
feat/native-contact-audit

Conversation

@Tarzelf

@Tarzelf Tarzelf commented Oct 18, 2026

Copy link
Copy Markdown
Contributor

Nominally 50 Hz replay observations can show a continuous grip while contact is interrupted between samples. Add passive 1 kHz contact instrumentation that replays frozen commands, records geometry and contact-frame force evidence, and requires exact agreement with every original sampled physical observation and complete initial/final outcome.

The audit verifies 14 frozen replays across two previously inspected scenes: 249,502 native ticks. A portable rerun using bundled inputs produces byte-identical compressed native traces and identical physical summaries for all 14. Baseline and smoothed students reach at most 75 ms and 89 ms of the required 100 ms lift dwell; compliance reaches 947 ms but still fails placement. These are diagnostic reproductions, not new policy trials or a controller improvement.

The change includes ten focused tests, immutable input/source hashes, retained failure evidence, a portable --from-audit path, compact result receipts and a sourced goal-image-to-actuator design note covering TypeSafe Jev. Jev has not been integrated or benchmarked. Complete input tapes and native traces are published with the research entry.

Validation: 10 focused tests passed, including an 80-native-tick exact-state probe. Independent source/result review found no high-impact methodological error. Primary and portable full runs each verified 14/14 episodes. Original physics, commands, force caps, success criteria and 18 s deadline are unchanged. CPU simulation remains the simplified rigid-box profile, separate from PiPER shoe geometry and hardware.

This PR is stacked on research/retention-baseline-20261016, which publishes the existing retention research separately.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant