Skip to content

Two cover mutants that only exact tangency kills, and two that nothing can - #52

Merged
d6g8k5htny-coder merged 4 commits into
chatgpt/drive-github-hardening-20260919from
claude/cover-tangency-on-hardening
Sep 24, 2026
Merged

d6g8k5htny-coder merged 4 commits into
chatgpt/drive-github-hardening-20260919from
claude/cover-tangency-on-hardening

Conversation

@d6g8k5htny-coder

Copy link
Copy Markdown
Owner

What this changes

Three controls in tests/test_cover.py (30, 31, 32) and the section of
research/cover/README.md that records what they establish. No code changes
— research/cover/regions.py is pinned and untouched.

Type of change

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

The finding

AnnulusBracketRegion.classify calls a cell provably disjoint from the annulus
when hi2 < rlo2 or lo2 > rhi2. Both comparisons are strict and each has a
mutant — <=, >= — differing from the shipped code only where a cell is
exactly tangent to a circle
. The reference covers never land on that equality,
so neither mutant dies by running the reference geometry. That is a fact about
the reference runs, not a reason to leave them alive.

Exact rationals make a tangent cell easy to build:

mutation witness live mutant caught by
hi2 < rlo2 → hi2 <= rlo2 [0, 3/50] × [0, 4/50], hi2 = 9/2500 + 16/2500 = 1/100 = r_lo² STRADDLE OUTSIDE control 30
lo2 > rhi2 → lo2 >= rhi2 [5, 6] × [0, 1], lo2 = 25 + 0 = 25 = r_hi² STRADDLE OUTSIDE control 31

Verified on scratch copies of regions.py — never by editing the library — and
each box kills exactly one:

live         c30=True  c31=True  c32=True
M_inner_le   c30=False c31=True  c32=True
M_outer_ge   c30=True  c31=False c32=True
M_axis_a     c30=True  c31=True  c32=True
M_axis_b     c30=True  c31=True  c32=True

OUTSIDE is the disposition whose rejection reason says proved disjoint, so a
cell touching the region and rejected as disjoint is how a cover loses area it
should have accounted for.

And two mutants that are provably equivalent, which is a different thing

The _axis_min_max operand flips look like two more of the same. They are not a
disjointness test at all:

lo = Fraction(0) if (a <= 0 <= b) else min(abs(a), abs(b))

a < 0 <= b can diverge only when a == 0, and the else branch then returns
min(|0|, |b|) = 0, which is what the if branch returns; a <= 0 < b can
diverge only when b == 0, and the else branch returns min(|a|, 0) = 0
likewise. The flips agree everywhere on a legal box, not merely at tangency
— the matrix above shows both surviving all three controls, as they must.

Control 32 therefore does not try to kill them. It asserts the equivalence
over a grid of more than seven hundred legal (a, b) pairs and against the
shipped helper, which is the only guard an equivalence claim can carry: if
_axis_min_max is later changed so the flips stop agreeing, the claim on the
README becomes false and control 32 fails.

A surviving mutant that is provably equivalent to the shipped code is not a hole
in the suite. One that survives because the reference geometry happens to miss
it is, and two of these were.

Why a port and not a merge

These controls exist on claude/drive-audit-github-migration-rrglpp, whose PR
was merged and then reverted (#32). research/cover/regions.py is byte-identical
on both lines
and is pinned by research/rn/candidates/inner_wedge_20260920_v1.json,
so only the controls and the README section move — no pinned byte is touched, and
the README section is written for the survivors this branch records rather than
a table it does not have. That follows the request on #38 to port rather than
rebase the whole branch.

Status discipline

  • No claim status was promoted — none was touched
  • Frozen bodies were not edited in place; regions.py, ledger.py, driver.py and __init__.py are all pinned and all unchanged
  • claims/graph.json unchanged
  • registers/json/work_events.json unchanged
  • 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 of anyone's work. Controls 30–32 are negative controls on this
repository's own test suite; no verdict is recorded and no independence credit
arises.

Verification

Every step of .github/workflows/ci.yml replayed locally — all 44, including
the mathematical replays and the complete pytest suite. CI_STEPS=44 CI_FAIL=0.

python3 -m pytest tests/test_cover.py -q      40 passed
python3 tools/mirror_quotes_check.py          problems=0
python3 -m pytest -q                          (full suite, in the replay)
  • All of the above pass locally

What this does not establish

Nothing about the cover's mathematics. Controls 30 and 31 say a tangent cell is
classified STRADDLE rather than OUTSIDE; control 32 says two rewrites of a
helper compute the same function. Both Pieces of D3-LEMMA-RN-UNIF remain
OPEN, no bound was tightened, and a classification is not a certificate.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY


Generated by Claude Code

d6g8k5htny-coder and others added 2 commits September 24, 2026 03:13
…g can

`AnnulusBracketRegion.classify` calls a cell provably disjoint from the annulus
when `hi2 < rlo2 or lo2 > rhi2`. Both comparisons are strict, and each has a
mutant -- `<=`, `>=` -- differing from the shipped code only where a cell is
exactly tangent to a circle. The reference covers never land on that equality,
so neither mutant dies by running the reference geometry. That is a fact about
the reference runs, not a reason to leave them alive.

Exact rationals make a tangent cell easy to build, and each of the two kills
exactly one mutant -- verified on scratch copies of `regions.py`, never by
editing the library, and never by touching bytes two certificates pin:

    [0, 3/50] x [0, 4/50]   hi2 = 9/2500 + 16/2500 = 1/100 = r_lo^2
    [5, 6]    x [0, 1]      lo2 = 25 + 0            = 25    = r_hi^2

    live        STRADDLE          mutant      OUTSIDE

`OUTSIDE` is the disposition whose rejection reason says *proved disjoint*, so a
cell touching the region and rejected as disjoint is how a cover loses area it
should have accounted for.

The `_axis_min_max` operand flips look like two more of the same and are not.
The helper is not a disjointness test:

    lo = Fraction(0) if (a <= 0 <= b) else min(abs(a), abs(b))

`a < 0 <= b` can diverge only when `a == 0`, and the else branch then returns
`min(|0|, |b|) = 0`, which is what the if branch returns; `a <= 0 < b` can
diverge only when `b == 0`, and the else branch returns `min(|a|, 0) = 0`
likewise. The flips agree everywhere on a legal box, not merely at tangency, so
nothing separates them. Control 32 does not try to kill them -- nothing can --
and asserts the equivalence over a grid of more than seven hundred legal pairs
and against the shipped helper, which is the only guard an equivalence claim can
carry. A surviving mutant that is provably equivalent to the shipped code is not
a hole in the suite; one that survives because the reference geometry happens to
miss it is, and two of these were.

Ported from the migration branch rather than merged from it: `regions.py` is
byte-identical on both lines and is pinned by
`research/rn/candidates/inner_wedge_20260920_v1.json`, so only the controls and
the README section move, and `research/cover/README.md` gets a section written
for the survivors this branch actually records rather than a table it does not
have.

What this does not establish: nothing about the cover's mathematics. Controls 30
and 31 say a tangent cell is classified STRADDLE rather than OUTSIDE; control 32
says two rewrites of a helper compute the same function. Both Pieces of
`D3-LEMMA-RN-UNIF` remain OPEN, no bound was tightened, and a classification is
not a certificate.

Co-Authored-By: Claude <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Cz7WZybv8znP64SpPj6sWY
@d6g8k5htny-coder
d6g8k5htny-coder marked this pull request as ready for review September 24, 2026 07:32
@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 8bd1f03 into chatgpt/drive-github-hardening-20260919 Sep 24, 2026
4 checks passed
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>
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.

1 participant