Repository navigation
Conversation
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Repeated camera/action learning runs spent far more time re-exporting sealed recordings than fitting the small CPU model. This preview adds an explicit
dvidia-train cachebuild/load path with independent manifest, protocol, split and source pins, preserving numeric learning inputs, labels and dataset identity. It leaves the frozen v0.1 controller and its 5/64 qualification result unchanged.On one registered 1,000-sample training episode, strict export took 1.636 seconds and three verified cache loads took 2.952, 2.831 and 3.214 ms. This is a warm shared-host data-preparation measurement, not a control or whole-study speed claim. Cache hits verify the derived snapshot; they do not reverify raw recordings.
A separate numerical preflight verifies the proposed reachable-reference action representation over 4,242 synthetic vectors and all 20 reference microsteps. Errors stay below the predeclared 1e-12 native-unit tolerance; 24 negative inputs are rejected. No models were fitted, no native simulations were run, and the proposed 164-attempt v0.2 screen remains unrun.
Validation: cache integrity/tamper, masks, correction-label authority, split isolation and CLI regression tests; all 19 registered runtime files remain byte-identical to the parent. The change adds no dependencies and preserves a 30 GiB disk reserve for cache construction. New cache tests run in the existing CPU CI matrix.
This PR is stacked on camera-to-action v0.1 (#32). The original study runtime remains pinned to c9a7c55; its publishing-only follow-up is d89b24c. This is a separate 0.1.3.dev5 preview.