[commdata2338]

Change and mathematical scope

Adds research/t74-source-partner-cut-audit/, a code-disjoint finite audit of
PR2530's eight new source-partner aliases and a scoped recovery-movement
certificate. Both frozen forward provider words put the intended original
dirty input in each new lender at handoff. Exact forward row replay and a
separate backward substitution agree on all sixteen observations, with 646
independent initial inputs per provider.

The seven fixed recovery blocks have 1,161 individual legal placements per
provider in their maximal connected intervals under adjacent commutation
and unchanged incident frames: 11, 9, 8, 10, 272, 406, 445. These are not a
joint schedule count or a global retiming optimum. Every crossed event and
both interval boundaries have explicit certificates. The source-68 left
boundary is algebraically movable but needs an extra frame climb; several
right boundaries change the actual four-dimensional subspace without changing
its dimension.

The certificate also records that older lenders 4 and 12 hold the mixtures
(x4-x68)/2 and x12+x31. The new-lender original-input invariant must not be
generalized to all older aliases.

This is a finite certificate and scoped proof, not a new helper inventory,
physical schedule, supplier or κ. Native inverse-transpose implementations,
Gaussian phases, final cleanup and target endpoint, buffer stock, full rank
invoice, owners, global/outer admission, precision and all-size/compiler
obligations are not recertified. The selected frontier is unchanged.

Task and overlap

The claim is in discussion #2606.
Full local task metadata and refresh evidence are in TASK.json.

Task card:

• Task ID / scope key: independent-moment / c191650c0fa8d2fa514e, refined to
  PR2530/eight-source-partners/chronological-lender-cuts/both-words/exact-rational-independent-audit.
• Base / exact predecessors: main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2;
  PR2530 b4012098048e12500b2d33d2e84be281915a27ac; its PR2462 supplier
  ef34071fee6dfb0286e9850a5cec4e36d9eb126d.
• Concrete question: do all eight new lenders hold their intended original
  inputs at their actual handoffs, and what fixed-block recovery movements
  preserve the complete B map without a frame change?
• Scope status / reason: refined. The card permits independent source-ledger
  reconstruction. The new lifetime implication and explicit cut certificates
  add coverage beyond another evaluation of an already checked moment root.
• Audited claim: PR2530's recovery-before-destructive-handoff assertion.
  The source construction is unchanged; the movement model is declared above.
• Method / domain: exact rationals and formal invertible operator words;
  both frozen forward providers, all eight named new lenders, seven fixed
  recovery blocks, nine recovery operations, explicit F_2^9 frame sets.
• Nearest overlap: PR2530 supplies the construction and full author checks;
  PR2571/PR2575 change owner packing; PR2588 changes helper labels/frame charges;
  PR2573's exhaustion excludes source-partner recovery; PR2315 audits an older
  borrow43 record and buffer accounting. Discussion Cost-aware source-partner helper assignments above PR2588 #2612 searches new source
  assignments. This package holds PR2530 fixed and audits its intermediate
  source values and recovery movement.
• Budget / observable improvement: audit only; no new cost or κ target.
  Require exact boundary witnesses and explicit semantic/frame failure controls.
• Success / falsification / cap: complete sixteen handoff identities and
  fourteen movement intervals, or a concrete failing cut. One bounded session;
  no unrestricted synthesis or complete supplier replay.
• Predecessor authors / artifact: zxv, Jeremy Kothe, Tekalign, Vlado
  Vojdanovski, Cinna Davis and retained lineage; the package above, reproduced
  with python3 -I -B research/t74-source-partner-cut-audit/verify.py.

At the final refresh, canonical main, priority revision 4, the selected record
and PR2530's exact head were unchanged. The final offline snapshot covered 812
open PRs at 16:22:35 UTC and 1,276 Discussion records at 16:27:10 UTC on
2026-10-15, both with complete pagination. The changed-file snapshot was from
15:59:04 UTC and flagged PR2571's file list as truncated beyond 100 files.
The later PR2621 timing/hosting claim was also screened as context, not audited
or used as a source. These searches are overlap evidence, not proof of novelty. Exact timestamp
and pins are in TASK.json.

Validation

On Linux, CPython 3.12.3, standard library only:

python3 -I -O -B research/t74-source-partner-cut-audit/verify.py
python3 -I -B research/t74-source-partner-cut-audit/test_verify.py
python3 -B scripts/research_package.py --package research/t74-source-partner-cut-audit --check-only
python3 -B scripts/research_package.py --package research/t74-source-partner-cut-audit

The registered replay runs the optimized verifier and focused test file. It
passes PASS_REGISTERED_RESEARCH_PACKAGE; all five tests pass (8.096 seconds
for the final focused file). The verifier prints:

PASS_T74_SOURCE_PARTNER_CUT_AUDIT providers=2 handoffs=16 recovery_gates=18 reverse_cuts=16 movement_windows=14 semantic_rejections=24 no_new_kappa

Validation includes exact input SHA-256/Git-blob pins, canonical certificate
comparison, noncommutative positive fixtures, 24 rejecting recovery/sign/order/
handoff controls, full 646-column B-map replay of all 28 interval endpoints,
and 28 immediately outside placements that fail map-plus-zero-climb
preservation. Two outside placements preserve the algebraic map and fail only
the frame condition. A target-only prelude corruption deliberately still passes
the lender-cut check and changes the full B map, documenting its coverage limit.

The predecessor's generator and checker were inspected but never imported or
executed. Normal replay does not write package files. No full native/global/
outer cold replay, formal-kernel proof, hosted CI or independent human review
is claimed.

Credit

PR2530 / zxv supplies the source-partner construction and original checks.
Jeremy Kothe / PR2462 supplies the combined T74 words and inherited pipeline;
Tekalign (@tekuuu) / PR2436 supplies lifted-target borrowing and transport;
Vlado Vojdanovski / PR2322 supplies coherent handoffs; Cinna Davis
(@1337isnot1337) / PR2226 supplies anticipatory dirty borrowing. The
PR1297/PR1319 and earlier supplier lineage retains its credit and Apache-2.0
license. This package adds the independent implementation, intermediate cut
certificates, bounded movement proof and tests.

Prepared for commdata2338 (@commdata2338), with substantial OpenAI Codex assistance.

AI assistance: submitted by @commdata2338 with substantial AI assistance. OpenAI Codex (GPT-6 Astra) did the derivation, the implementation, the checking and the writing; it was packaged and submitted with Claude Code. No independent human review is claimed.

Open next questions

The README and TASK.json give two bounded follow-ups: bind prescribed
invertible-mixture lender rows to their complete compensation/cleanup contracts,
starting with older roles 4 and 12; or supply exact inverse rows and an available
frame for one three-source recovery candidate before attempting a new word.
Neither source availability nor an algebraically movable block establishes a
favorable native bill or supplier admission.

Workflow feedback (optional)

The intermediate-cut audit exposed a useful distinction between algebraic
movement and frame-preserving movement without launching a full record replay.
Future recovery constructions could publish a small table of restored rows,
the last required reads and the first destructive handoffs beside their endpoint
receipts. That would make the next independent review more direct.

🤖 Generated with Claude Code

https://claude.ai/code/session_01X86sCSXdoDd61SDFfC1PKG