Repository navigation
Conversation
* Add independent camera and drive checks for goal verification * Align camera evidence and speed up CPU image checks * Keep drive summaries compact and evidence-referenced * Publish sensor corroboration v0.4 benchmark evidence * Resolve publisher replay directory before report regeneration * Tolerate cross-runtime replay rounding * Rebuild publication export integrity seal * Chunk Hugging Face evidence uploads * Use Hub Xet for binary evidence assets * Resume Xet-backed sensor evidence publication --------- Co-authored-by: DVIDIA <hello@dvidia.org>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The prior Jev telemetry evaluation recognized only 1/8 completed scenes. Response forensics show three visible-scene abstentions and four covered-scene incomplete verdicts; request payloads and parsing were correct. This adds a local CPU verifier that checks a trusted goal against fresh, typed, uncertainty-bounded physical observations and explains complete, incomplete or unknown decisions.
A frozen 200-case native benchmark recognizes 16/16 completed scenes with 0/48 native false completions. All 128 evidence faults produce their specified conservative decisions. Eight self-consistent sensor forgeries still cause eight physical false completions; they remain visible in separate and aggregate metrics (8/56 overall negatives). Native verifier compute p95 is 0.317 ms, excluding sensing and the required 200 ms history. This is measured-state verification, not learned perception or a robot controller.
This PR is stacked on #23 to keep the earlier frozen Jev protocol separate. Source frozen before evaluation: b4c76b1. Results, forensics, protocol and independent audit: benchmarks/local-goal-verifier-20261019/README.md.
A wheel smoke test at publication source 180e26c passed in a clean environment without NumPy or MuJoCo: CLI help and both complete/moving fixtures work outside the source checkout. Windows matrix setup is currently blocked by Chocolatey returning HTTP 503 for FFmpeg, before test execution; verifier offline CI passes.