Skip to content

Independent forward replay and mutation controls for words 7000 and 6030 - #7540

Open
ahrazzle wants to merge 1 commit into
CrocSwap:mainfrom
ahrazzle:independent-global-rank-replay
Open

ahrazzle wants to merge 1 commit into
CrocSwap:mainfrom
ahrazzle:independent-global-rank-replay

Conversation

@ahrazzle

Copy link
Copy Markdown
Contributor

Change and mathematical scope

Status: audit tooling. Adds research/independent-global-rank-replay/ and tests/test_independent_global_rank_replay.py. No new construction, kappa or supplier, and no selected result changes.

The script replays the native words of #7000 and #6030 forward over exact fractions. It rebuilds the paid and target profiles from the replay, finds the crossing with one solver, and proves the root is unique on [0, 0.01] with a derivative bound. The moments change sign 15 and 19 times when ordered by rank, so Descartes' rule would not give uniqueness.

Results at the base commit:

Item Value
word 7000 credit mass / paid mass per call 7650 / 10962
word 7000 rank surplus 1520
word 7000 crossing 0.0036300581746883670859, 4.6e-12 above the tested 0.00363005817007
word 6030 productive / paid / surplus 13444 / 66834 / 760
word 6030 root 0.0031258115579118972, inside (0.00312581155791, 0.00312581155792)

For word 6030 the rebuilt T and P equal the record exactly, with the query term (15 query shears times 104 ports on P[1]). The solver sees only the rebuilt profiles.

Independent of the audited packages: the forward replay, the profile rebuild, the word 6030 known-answer test and the mutation controls. Not independent: the crossing uses the repository's own invoice formula, so matching the word 7000 record is a reproduction. Caller tables are read from the records, and the main-row block comes from the word's own declared parameters.

Not covered: kappa, the supplier, the caller identity and the outer gate. Split tails are not identified from the word (helper 450 has two candidates with equal frames and rank).

Task and overlap

Base commit e54523517d0f03e58352e647c403a5af3e525a57. Task card independent-moment, scope refined to two named words (see TASK.json).

Audited packages: word 7000 (research/shared-wire-split-composition) and word 6030 (research/p104-helper470-twin-rung-removal). Both are read as data and no code is imported. Added value: a known-answer test that never reads the recorded T, P or answer into the solver, and controls that edit the word itself.

Validation

python3 -I -B research/independent-global-rank-replay/replay.py
python3 -I -B tests/test_independent_global_rank_replay.py
python3.12 -I -B scripts/research_package.py --package research/independent-global-rank-replay

Expected: the first prints PASS after the JSON (about 40 seconds, cap 300 seconds in VERIFY.json). The second reports 14 tests OK (about 35 seconds, same cap). The third is the repository package validator. It exits 0 and ends with PASS_REGISTERED_RESEARCH_PACKAGE research/independent-global-rank-replay outcome=completed, after running the replay and the 14 tests. It needs Python 3.12. Run on CPython 3.12, 3.13 and 3.14, and with -O. Python 3.11 was not run. Network off, writes confined, stdlib only.

Negative controls, each rejected by its target check while the unmutated word passes it:

  • ADD 1478 += 759 inserted at event 40637: double-credit check (a helper row is no longer the identity and an input is shared).
  • Sign flip of one main-row ADD: main-row check.
  • An ADD with scalar 0: nonzero-scalar check.
  • A public MOVE down and back up: direction check.
  • A deleted public MOVE: frame-chain check.
  • word 6030 rebuilt without the query term: P equals record.
  • One rank-8 credit removed per call from word 7000: root at least the tested value.

Not run: the repository's full test suite and the other research replays.

Credit

Audited work is by the authors of the two packages and their predecessors, credited in those packages. An independent reviewer found weaknesses in an earlier version of these checks, and this version fixes them. Written and run by Claude models (Anthropic) under human oversight. Apache-2.0.

Open next questions

  • Check the main-row block against the symbolic caller identity in data/caller.json instead of the word's own declaration. No blocker.
  • Identify split tails from the word, starting with the two candidates for helper 450. No blocker.

🤖 Generated with Claude Code

Automated posting by agentic team with human oversight.

Forward replay, invoice rebuild, unique-root crossing check, a known-answer
test on 6030 and seven mutation controls. Audit tooling only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.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