Repository navigation
Qualify native engineering convergence with evidence checks and a Claude child repair - #5
Merged
seathatflowsinourveins merged 5 commits intoSep 20, 2026
Conversation
seathatflowsinourveins
pushed a commit
that referenced
this pull request
Sep 24, 2026
…ix mover-v3 step-down text Fix round for audit gap #5 review findings. - scripts/trading_gates.py: new source_matches condition (standalone or as an all_of member). The receipt names an in-tree file by path and sha256; the file's bytes must hash to that value, and each [receipt, source] pointer pair must be equal with the same JSON types. Paths that are absolute, contain '..' or a backslash, or resolve outside the tree are refused. - catalogs/us-equities/gates-20260922.json: each leverage-ladder-1x/2x/4x all_of now also requires /schema_version == 1, /kind == "leverage_ladder_rung_receipt", /rung, /source/certified_run_status == "passed", and a source_matches binding. The binding ties the receipt's /leverage and certified_run_status to the /leverage and /status of the committed paper-output.json it names. The preregistered receipt schema in the notes gains source.{paper_output_path, paper_output_sha256, certified_run_status}. The notes also state a limit: needs_attention is a count across runs, is not bound to the hashed file, and is checked at the manual qualification. - mover-v3 README: the step-down list no longer calls seconds_above_next_lower_rung_ceiling a breach. It names achieved leverage above the ceiling in force as a proposed v3 rule (runner.py has no step-down on it today). seconds_above is described as a promotion input. - README-safety.md: documents the binding and its limits. - Tests: source_matches shape and behaviour. Every ladder rung rejects a receipt whose leverage block or status was edited, whose sha256 is wrong, or whose source is missing, outside the tree or absent. A rung also rejects a receipt that stitches one run's leverage onto another run's passed status, and one whose kind, schema_version or rung is wrong. Trials that never pass the lower cap still fail only on the achievement members. AchievedLeverageGateTraceTests now builds a hashed synthetic source. source-hashes.json and manifests/evidence.json entries are refreshed for the changed files only. Offline unit tests on synthetic receipts and sources only. No paper trial was run, and no rung receipt exists. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
seathatflowsinourveins
deleted the
codex/native-convergence-2026-09-20
branch
September 25, 2026 18:47
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 public convergence protocol had no executable guard against contradictory adoption or savings claims, and clean WSL Claude authentication had not proved a real isolated child repair. This adds a general engineering entrypoint, scoped evidence validation, an observed native Opus 5 child repair, and checksum-pinned offline workflow security analysis.
Validation:404 local standard-library-profile tests pass with47 explicit optional-profile skips; all35 new focused checks execute. Publication/catalog/experiment validators, explorer reproducibility and redacted secret scan pass. A preceding partial private Python profile produced10 missing-pytz errors and is retained in the validation record. This is not an optional-SDK/GPU/all-host qualification, provider-cancellation proof, or universal SOTA claim.