Repository navigation
Verify partial-evidence scoring against actual Inspect SQLite tools - #5
Merged
Merged
Conversation
YusefSyed
marked this pull request as ready for review
October 3, 2026 07:20
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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 partial-evidence scorer previously had synthetic evidence but no native Inspect execution path. Add a separate integration that runs the existing 13 SQLite tool fixtures in restricted Docker sandboxes, captures private producer-side receipts, seals scores before reading the independent database oracle, and verifies uncertainty when receipts, effect records or entire actions are withheld.
The scorer receives no privileged database history, fixture labels or tool/model prose. Two fresh scripted runs must reproduce normalized observations, predictions, reports and database bytes. Native logs and source hashes support offline verification; this is finite scripted integration evidence, not live-model or production validation.
Validation: 73 bridge tests, 40 legacy tests, Ruff, strict projection mypy, all 17 frozen core artifacts and the unchanged actual-tool baseline passed in CI. All 13 new Docker scenarios ran twice, and the downloaded raw evidence passed local offline verification. The normalized observations, predictions, reports and DB bytes match across runs. Core Python 3.12/3.13/3.14 checks passed. Versioned evidence and reproduction instructions are in artifacts/inspect-v3-bridge/v1/REPORT.md.
Each execution yields five related evidence views: 21 binary decisions agree with the independent DB oracle and 44 abstain per run. These are not independent model samples. The committed evidence verifier is also required in CI. Core, legacy registry and prior frozen artifacts remain unchanged.