Repository navigation
Which files a certificate binds by digest was learnable only by breaking one - #51
Merged
d6g8k5htny-coder merged 7 commits intoSep 24, 2026
Conversation
…ing one
Eight candidate records under `research/` bind repository content by SHA-256 as
source identity. The replay checkers that consume them refuse the tree when a
byte has moved, which is the right discipline. What was missing is that no
document said which files those are, and the rejection a replay prints names
the dependency but not the certificate:
tools/twelve_project_check.py REJECTED
tools/h3_rn_n6_check.py REJECTED: repository dependency mismatch:
quarantine/EXCLUSIONS.json
Three separate changes paid a cycle for that in four days --
`research/bands/ladder.py`, `research/cover/ledger.py` and
`quarantine/EXCLUSIONS.json`.
`research/PINNED_SOURCES.md` is now generated from the certificates themselves
and verified against the tree on every CI run. A mismatch is reported once, by
file, with the certificate that names it, both digests, and the remedy -- a
re-pin on the lane that owns the certificate, never an edit to the expected
digest.
certificates=8 containers=15 pinned_files=34 pinned_archive_members=6
sites=106 compared=40 digest_matches=40 unresolved=0
keys_not_repository_relative=47 problems=0
Three of the thirty-four are checkers in `tools/` -- `drive_coverage.py`,
`manifest_integrity_check.py`, `hermite_envelope_report.py` -- and one is
`docs/HERMITE_GAUSSIAN_ENVELOPE.md`. Editing a checker can break a certificate,
and nothing in the tree said so.
## Three shapes, and one real ambiguity
A tool that knows one shape under-reports the surface. The certificates use
`path -> {bytes, sha256}`, `path -> "<sha256>"` (a bare string, which is how the
H3 candidate binds `ladder.py`), and `<archive>::<member>` for content sealed in
a ZIP carrier. All three are read, and a member is hashed out of its carrier.
The ambiguity is the root, not the shape. The recovered Lean bundle keys its
`members` inside the bundle and one is called `README.md`; resolving it here
compares a file the certificate never named. That is the defect this repository
already found and fixed once in its seal checker, which "resolved each declared
path by its last component, so a line naming a file in a subdirectory was
compared against a same-named sibling".
`classify()` settles it from evidence -- a container is repository-relative when
at least one key resolves here to the bytes it pins -- and then the generated
index *records* that classification, and later runs use what was recorded rather
than re-deriving it. Deriving every time leaves a hole: a container whose every
pin drifted loses the evidence, reclassifies itself out of the check, and
reports a clean run over changed files. A control builds exactly that tree and
requires both files to be named.
The vacuity floor tests comparisons made, not comparisons that matched. A run in
which every binding was read and every one differed is a loud failure, not a
vacuous one, and calling it vacuous would bury the mismatches it just found.
## Controls
`tests/test_pinned_sources.py` holds seventeen, every one running the tool
through its CLI against a synthetic root. Ten weakened copies were built in a
scratchpad and run against them: the bare-digest shape ignored, archive
resolution removed, the recorded classification ignored, the mismatch report
silenced, the byte-count check removed, the vacuity floor removed, the index
comparison removed, the new-container check removed, the classification-change
check removed, and every container forced to repository-relative. Each was
caught by the control written for it; the last is caught on the real tree, by
the `README.md` false mismatch it creates.
`CLAUDE.md` carries the rule and `AGENTS.md` one line pointing at the page,
which is the whole point of having it. That line puts AGENTS.md at exactly the
45 lines `tests/test_agents_md_points_and_does_not_legislate` allows, so the
next addition there has to replace it or move the cap -- flagged rather than
left as a trap.
The CI step needed two edits, not one: `tools/run_checks.py` keeps a deliberate
duplicate of the reviewed command set and `tools/closure_pipeline.py check-plan`
refused the workflow until the new command was added there too --
FAIL closure_pipeline: unsupported or dynamic CI command:
'python tools/pinned_sources_check.py'
-- which is that allowlist doing exactly what its comment says it is for.
## What this does not establish
A pin is an identity statement and nothing more: these are the bytes the
certificate was produced against. Nothing here says whether a certificate is
correct, whether its mathematics holds, or whether a bound file is current or
authoritative. No claim, premise, obligation or gate is discharged, and no
status moves. Forty bindings verifying is forty digests matching.
Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY
…ces-inventory The base gained the vault path map (#44) and the checked inner-wedge accounting (#48). No textual conflict: both sides appended to tools/run_checks.py's REQUIRED_COMMANDS and to the CI step list, and git took both. The certificate inventory is unchanged -- 8 certificates, 34 files and 6 archive members, all 40 digests still matching -- because neither merged PR touched a pinned byte. There was a semantic one. `AGENTS.md` is capped at 45 lines by `test_agents_md_points_and_does_not_legislate`, both sides added a pointer, and the merged file was 46. The base's pointer to `quarantine/PATHS.md` is the one that stays: mine is dropped, because the rule it pointed at is in `CLAUDE.md`, which AGENTS.md's first paragraph already tells every agent to read. Raising that cap is a decision about someone else's file and not one to take in a merge. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY
…nned-sources-inventory
This was referenced Sep 24, 2026
…lker Certificates nest source identity as siblings in a list, not only as map keys, bare digests, or archive members. research/parallel/lpw/candidate.json binds three drive-mirror paths that way; their digests match this tree and the inventory omitted them. The same shape is how DEPENDENCIES.json names its imports. The generated index now lists those three paths. A synthetic control fails if the walker drops the list, and the real-tree test requires the three paths to be present. certificates=10 containers=19 pinned_files=37 pinned_archive_members=6 sites=128 compared=43 digest_matches=43 unresolved=0 keys_not_repository_relative=56 problems=0 Co-authored-by: Dylan Roy <d6g8k5htny@privaterelay.appleid.com>
…nned-sources-inventory
…nned-sources-inventory
d6g8k5htny-coder
marked this pull request as ready for review
September 24, 2026 11:26
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
…nned-sources-inventory
8 of 14 tasks
d6g8k5htny-coder
merged commit Sep 24, 2026
b89448d
into
chatgpt/drive-github-hardening-20260919
4 checks passed
cursor Bot
pushed a commit
that referenced
this pull request
Sep 24, 2026
The hardening tip observed at this edit moves from 8bd1f03 to LOCK b89448d after #51. The prior tip stays an earlier observation that was not re-executed. Probe runners were not re-executed. Honesty labels stay REFUSED/EMPTY/ABSENT/PARTIAL/REFUSED_NOT_24JET. PACKET retargets only the STATUS_JETMOD transcription digest. OBL-H5-JETMOD stays OPEN. Co-authored-by: Dylan Roy <d6g8k5htny@privaterelay.appleid.com>
9 of 15 tasks
d6g8k5htny-coder
added a commit
that referenced
this pull request
Sep 24, 2026
* docs: advance inventable tip observation to 8bd1f03 The hardening tip observed at this edit moves from c82c935 to LOCK 8bd1f03 after #52. Probe runners were not re-executed. Honesty labels stay REFUSED/EMPTY/ABSENT/PARTIAL/REFUSED_NOT_24JET. PACKET retargets only the STATUS_JETMOD transcription digest. OBL-H5-JETMOD stays OPEN. Co-authored-by: Dylan Roy <d6g8k5htny@privaterelay.appleid.com> * docs: advance inventable tip observation to b89448d The hardening tip observed at this edit moves from 8bd1f03 to LOCK b89448d after #51. The prior tip stays an earlier observation that was not re-executed. Probe runners were not re-executed. Honesty labels stay REFUSED/EMPTY/ABSENT/PARTIAL/REFUSED_NOT_24JET. PACKET retargets only the STATUS_JETMOD transcription digest. OBL-H5-JETMOD stays OPEN. Co-authored-by: Dylan Roy <d6g8k5htny@privaterelay.appleid.com> --------- Co-authored-by: Cursor Agent <cursoragent@cursor.com>
This was referenced Nov 9, 2026
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.
What this changes
New
tools/pinned_sources_check.pyand the page it generates,research/PINNED_SOURCES.md: every repository file that a certificate underresearch/binds by SHA-256 as source identity, verified against the tree onevery CI run.
tests/test_pinned_sources.pyis new.CLAUDE.mdcarries therule,
AGENTS.mdone line pointing at the page, and.github/workflows/ci.ymlplus the duplicate command allowlist in
tools/run_checks.pycarry the newstep. No pinned byte is touched.
Type of change
The problem
Eight candidate records bind repository content by digest. The replay checkers
that consume them refuse the tree when a byte moves — correct, and untouched
here. What was missing is that nothing said which files those are, and the
rejection a replay prints names the dependency but not the certificate:
Three separate changes paid a cycle for that in four days —
research/bands/ladder.py,research/cover/ledger.pyandquarantine/EXCLUSIONS.json.Three of the thirty-four are checkers in
tools/—drive_coverage.py,manifest_integrity_check.py,hermite_envelope_report.py— and one isdocs/HERMITE_GAUSSIAN_ENVELOPE.md. Editing a checker can break a certificate,and nothing in the tree said so.
This supersedes the two inventories I posted on #38. Both were produced by
ad-hoc scripts and both were wrong, in both directions; the generated page is
the count now.
Three shapes, and one real ambiguity
A tool that knows one shape under-reports the surface:
path → {bytes, sha256}path → "<sha256>"h3/candidate.jsonbindsladder.py<archive>::<member>The ambiguity is the root, not the shape. The recovered Lean bundle keys its
membersinside the bundle and one is calledREADME.md; resolving it at therepository root compares a file the certificate never named. That is the defect
this repository already found and fixed once in its seal checker, which
"resolved each declared path by its last component, so a line naming a file in
a subdirectory was compared against a same-named sibling".
classify()settles it from evidence — a container is repository-relative whenat least one key resolves here to the bytes it pins — and then the generated
index records that classification, and later runs use what was recorded.
Re-deriving every time leaves a hole: a container whose every pin drifted
loses the evidence, reclassifies itself out of the check, and reports a clean
run over changed files. A control builds exactly that tree and requires both
files to be named.
The vacuity floor tests comparisons made, not comparisons that matched. A run
in which every binding was read and every one differed is a loud failure, not a
vacuous one, and calling it vacuous would bury the mismatches it just found.
Status discipline
PINNED_SOURCES.mdis modified hereclaims/graph.jsonunchangedregisters/json/work_events.jsonunchanged — neither a claim nor a publicationIf this is a review
Not a review. The tool reads certificates and compares digests; it forms no
verdict on any certificate and no independence credit arises.
Controls
tests/test_pinned_sources.py— seventeen, every one running the tool throughits CLI against a synthetic root. Ten weakened copies were built in a scratchpad
— never by editing the library — and run against them:
…bare_digest_shape_is_recognised, real tree…resolved_inside_its_carrier, real tree…wrong_byte_count_is_refused…edited_index_is_refused_and_the_row_is_named…must_be_declared_by_regenerating…every_pin_drifted_is_still_comparedREADME.mdfalse mismatchEach was caught by the control written for it.
Verification
Every step of
.github/workflows/ci.ymlreplayed locally — all 45, includingthe mathematical replays and the complete pytest suite.
CI_STEPS=45 CI_FAIL=0.Two things the replay caught that a push would otherwise have discovered in CI:
tools/closure_pipeline.py check-planrefused the workflow until the newcommand was added to the deliberate duplicate allowlist in
tools/run_checks.py—unsupported or dynamic CI command: 'python tools/pinned_sources_check.py'— which is that allowlist doing exactly whatits own comment says it is for.
tests/test_bridge.py::test_agents_md_points_and_does_not_legislatecapsAGENTS.mdat 45 lines and my first pointer was nine. The detail moved toCLAUDE.mdand one line stayed. That line putsAGENTS.mdat exactly 45,so the next addition there must replace it or move the cap — flagging that
rather than leaving a trap.
What this does not establish
A pin is an identity statement and nothing more: these are the bytes the
certificate was produced against. Nothing here says whether a certificate is
correct, whether its mathematics holds, or whether a bound file is current or
authoritative. Forty bindings verifying is forty digests matching. No claim,
premise, obligation or gate is discharged, and no status moves.
🤖 Generated with Claude Code
https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY
Generated by Claude Code