Skip to content

The quarantine summary counted twenty-two comparisons where sixteen ran - #50

Merged
d6g8k5htny-coder merged 2 commits into
chatgpt/drive-github-hardening-20260919from
claude/quarantine-digest-coverage
Sep 24, 2026
Merged

d6g8k5htny-coder merged 2 commits into
chatgpt/drive-github-hardening-20260919from
claude/quarantine-digest-coverage

Conversation

@d6g8k5htny-coder

Copy link
Copy Markdown
Owner

What this changes

tools/quarantine_check.py now reports how many exclusions invariant 3 actually compared, refuses a record it cannot compare that does not say why, and refuses a run in which it compared nothing. quarantine/README.md records the split; registers/CONSUMERS.json records the new test as a machine consumer of the quarantine_index tab; tests/test_quarantine_digest_coverage.py is new. quarantine/EXCLUSIONS.json is byte-identical to its parent commit — see Where the reasons live below.

Type of change

  • Research content (a new or amended mathematical object)
  • Review verdict
  • Register update
  • Governance / protocol
  • Tooling, tests or CI
  • Documentation only (quarantine/README.md, alongside the tooling)

The defect

Invariant 3 is "no excluded payload digest appears in any repository manifest". It can only ask that of a record carrying a payload_sha256. Six of the twenty-two exclusions do not carry one, so the loop skipped them in silence while the summary line printed exclusions=22. Anyone auditing quarantine coverage off that line counted twenty-two comparisons where sixteen had been made.

before   exclusions=22 archive_members=14 manifest_digests=2965 …
after    exclusions=22 digest_comparable=16 digest_not_compared=6 archive_members=14 …

This is an accounting fix. The same sixteen digests are compared as before; nothing here compares more.

Why not simply widen the invariant

Because it would be wrong, not merely awkward. Q-R17-DUP-001 is an EXACT_DUPLICATE whose digest does exist, in a free-text identity field — and an exact duplicate shares its bytes with a retained keeper by definition. Promoting that digest into payload_sha256 would fire invariant 3 on the keeper and break the build on a correct tree.

Verified 2026-09-24, read off the manifests rather than asserted:

Drive id manifest row
excluded surplus copy 1Y_3zFonLsFIAHP5KSkUfsJXqHZXIUAL2 tree-only DO_NOT_PORT, stored: false, sha256: null
retained keeper 1Hc8dJvdh504xKBBHXswU5Ly-_8uYp_Sv stored, under the 2026-09-15 KIMI FINAL INTAKE lane

Nothing excluded has leaked, and two controls re-derive that from the manifests on every run.

The other five are three EXISTING_CONTAINER folder pointers (one of them the 99_DO_NOT_OPEN vault), a native Google document for which the corpus declares no payload digest, and a SUPERSEDED record identified by byte count alone.

What the checker now does

  • counts and reports digest_comparable and digest_not_compared separately;
  • refuses a record it cannot compare that does not say why, so a new exclusion cannot join the unchecked set by omitting a field;
  • refuses a run in which nothing was compared at all — the vacuity floor verify_manifests.py already carries;
  • refuses a declared reason that has stopped being true — one naming a key that is not an exclusion, or one whose record has since gained a digest.

Where the reasons live, and why that is not where they belong

research/rn/candidates/inner_wedge_20260920_v1.json pins quarantine/EXCLUSIONS.json twice as source identity, at 19,555 bytes / 8a5a8901…. Adding a field there fails four replay checkers closed — found the way everyone else finds it, by doing it:

tools/twelve_project_check.py      REJECTED
tools/h3_rn_n6_check.py            REJECTED: repository dependency mismatch: quarantine/EXCLUSIONS.json
tools/rn_inner_wedge_check.py      REJECTED: repository dependency mismatch: quarantine/EXCLUSIONS.json
tools/rn_bernstein_sharp_check.py  REJECTED: repository dependency mismatch: quarantine/EXCLUSIONS.json

A label is not worth breaking a pin for, so these six reasons are declared in DIGEST_NOT_COMPARED in the checker instead. A new exclusion still states its reason on the record, in a payload_digest_not_compared field; both routes satisfy the check and both are controlled.

Two further controls assert the pin itself and read the expected digest back out of the certificate, so the constants and the binding cannot drift apart, and the next person to reach for that file learns about it here rather than from a red CI run. That discoverability gap is wider than this one file — twelve certificates pin thirty-eight paths between them and no document in the tree mentions them — and I would like to close it separately rather than widen this PR.

Status discipline

  • No claim status was promoted — none was touched. No exclusion was restored, reclassified or discharged
  • Frozen bodies were not edited in place; quarantine/EXCLUSIONS.json is byte-identical to its parent commit
  • claims/graph.json unchanged — no claim or premise status moved
  • registers/json/work_events.json unchanged — this is neither a claim nor a publication
  • Nothing composes the 2D track with the 3D lifetime track
  • Nothing moves the prize track into the q0 dependency graph

If this is a review

Not a review. No verdict is recorded and no independence credit arises.

Controls

tests/test_quarantine_digest_coverage.py — twenty-one controls, every one invoking the checker through its CLI against synthetic or mutated inputs, so a path bound at import time cannot silently re-check the real repository.

Nine weakened copies of the checker were built in a scratchpad — never by editing the library — and run against the controls:

weakening caught by
reason requirement removed …must_say_why, …blank_reason…, …cannot_join_the_unchecked_set_silently
whitespace accepted as a reason test_a_blank_reason_is_not_a_reason
vacuity floor removed test_a_run_that_compared_nothing_is_refused
reason turned into an opt-out from the comparison test_a_reason_does_not_excuse_a_record_that_does_carry_a_digest
leak report silenced test_a_leaked_payload_is_still_caught
the two counters swapped six controls
stale-table "record gained a digest" arm removed …gained_a_digest_is_refused
stale-table "key is gone" arm removed …that_is_gone_is_refused
reason lookup short-circuited to always satisfied three controls

Each was caught, by the control written for it and by no other.

Verification

Every step of .github/workflows/ci.yml was replayed locally against this tree — all 44, including the mathematical replays and the complete pytest suite.

python3 tools/registers_import.py --check
python3 tools/registers_check.py
python3 tools/claims_check.py
python3 tools/quarantine_check.py
python3 tools/consumers_check.py
python3 tools/verify_manifests.py
python3 -m pytest -q
  • All of the above pass locally

What this does not establish

Nothing about any claim, premise, obligation, gate or theorem. A pass says sixteen excluded digests are absent from every manifest in this tree, and that six records sit outside that comparison for stated reasons — not that those six are harmless, not that their contents were reviewed, and not that the logical quarantine is complete. Counting what a checker skipped is not checking it.

Base

Targeted at chatgpt/drive-github-hardening-20260919 rather than claude/drive-audit-github-migration-rrglpp: PR #2 was merged and then reverted by #32, PR #3 is closed NEVER-MERGE, and main's landing page names this branch as the active research line. The three substantive files were byte-identical on both branches when this was written.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY


Generated by Claude Code

Invariant 3 of `tools/quarantine_check.py` is "no excluded payload digest
appears in any repository manifest". It can only ask that of a record carrying
a `payload_sha256`. Six of the twenty-two exclusions do not carry one, so the
loop skipped them in silence while the summary line printed `exclusions=22`.
A reader auditing quarantine coverage off that line counted twenty-two checks
where sixteen had been performed.

This is an accounting fix. The same sixteen digests are compared as before,
and nothing here compares more.

Widening the invariant to all twenty-two would be wrong rather than merely
awkward. `Q-R17-DUP-001` is an EXACT_DUPLICATE whose digest does exist, in a
free-text `identity` field; an exact duplicate shares its bytes with a
*retained keeper* by definition, so promoting that digest into
`payload_sha256` would fire invariant 3 on the keeper and break the build on a
correct tree. Verified: the excluded surplus copy
(1Y_3zFonLsFIAHP5KSkUfsJXqHZXIUAL2) is tree-only DO_NOT_PORT with `sha256:
null`, and the one stored row carrying those bytes is the keeper, a different
Drive object (1Hc8dJvdh504xKBBHXswU5Ly-_8uYp_Sv) under the 2026-09-15 KIMI
FINAL INTAKE lane. Nothing excluded has leaked, and two controls read that off
the manifests rather than asserting it.

What the checker now does:

  - counts and reports `digest_comparable=16 digest_not_compared=6`;
  - refuses a record it cannot compare that does not say why, so a new
    exclusion cannot join the unchecked set by omitting a field;
  - refuses a run in which nothing was compared at all, the same vacuity floor
    `verify_manifests.py` already carries;
  - refuses a declared reason that has stopped being true -- one naming a key
    that is not an exclusion, or one whose record has since gained a digest.

The reasons live in `DIGEST_NOT_COMPARED` in the checker rather than on the
records because `research/rn/candidates/inner_wedge_20260920_v1.json` pins
`quarantine/EXCLUSIONS.json` twice as source identity, at 19,555 bytes /
`8a5a8901...`; adding a field there fails four replay checkers closed. A label
is not worth breaking a pin for. A new exclusion still states its reason on the
record, in a `payload_digest_not_compared` field, and both routes are
controlled. Two further controls assert the pin itself and read the expected
digest back out of the certificate, so the next person to reach for that file
learns about the binding here instead of from a red CI run.

`tests/test_quarantine_digest_coverage.py` holds twenty-one controls, every one
invoking the checker through its CLI against synthetic or mutated inputs so a
path bound at import time cannot silently re-check the real repository. Nine
weakened copies of the checker were built in the scratchpad and run against
them -- reason requirement removed, whitespace accepted as a reason, vacuity
floor removed, the reason turned into an opt-out from the comparison, the leak
report silenced, the two counters swapped, each arm of the stale-table guard
removed, and the reason lookup short-circuited to always satisfied. Each was
caught, by the control written for it and by no other.

What this does not establish: nothing about any claim, premise, obligation,
gate or theorem. It moves no status and clears nothing. A pass says sixteen
excluded digests are absent from every manifest in this tree and that six
records are outside that comparison for stated reasons -- not that those six
are harmless, not that their contents were reviewed, and not that the
quarantine is complete.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY
Keep the digest-coverage accounting on the checker, its tests, the
CONSUMERS registration, and the README. Place the tip path-map section
after that accounting. quarantine/EXCLUSIONS.json stays the tip pin.

Co-authored-by: Dylan Roy <d6g8k5htny@privaterelay.appleid.com>
@d6g8k5htny-coder
d6g8k5htny-coder marked this pull request as ready for review September 24, 2026 03:45
@chatgpt-codex-connector

Copy link
Copy Markdown

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

@d6g8k5htny-coder
d6g8k5htny-coder merged commit c82c935 into chatgpt/drive-github-hardening-20260919 Sep 24, 2026
4 checks passed
cursor Bot pushed a commit that referenced this pull request Sep 24, 2026
…thub-chatgpt-xxw5kp

Tip-align onto c82c935 (quarantine comparison count, #50). Prior tip ac33581 remains an ancestor. Draft PR #21 base stays chatgpt/drive-github-hardening-20260919.

Co-authored-by: Dylan Roy <d6g8k5htny@privaterelay.appleid.com>
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.

2 participants