[tekuuu]

Change and mathematical scope

PR2995 leaves open whether its restoration-preserving F0 scalar fault survives the actual native operators. This mathematical audit closes that binding on the stated witness family. The same allowed completed ADD defect produces 4 u T maximum data error while all 480 physical helpers restore and all eight private fields return zero. The scalar witness and reserve remain credited to PR2995; the new result is their exact native realization and restricted tensor lift.

The proof tracks the constant input through its nonconstant intermediate frame vectors, checks the literal child/adapter recipes including global phase, and handles residual boundaries without assuming arbitrary helpers return to their original physical bits. This rules out a data-only accuracy guarantee from the permissive relative ADD specification plus restoration. It does not exhibit failure of a concrete rounding algorithm: exact grid-preserving ADDs can avoid the fault. No supplier or new κ is claimed.

Task and overlap

Scientific base 836b19bfa309ceafbc8ee32a15e8c4b10dde7e11; synchronized submission base d8e0f83f00191df62174b641e9fff055c357f641. The coordination update changes no selected result or live input. Predecessors: PR2995 8ba0358b62edbeb16a0842473500c468d8c0759b, PR2965 526f2c242202b31b15e09a7502dfb8ff757b2bad, and the pinned joint-query/native archive. Exact dependency/member pins and task scope are in research/qq16-native-dirty-fault/{SOURCES,TASK}.json.

This takes PR2995's released native-binding handoff. Canonical claim. The adversarial audit card is changed from the already-covered scatter obligation to this producer implication. Updated PR metadata, issues and canonical comments were refreshed before submission; the newer V45 candidate and invoice audits concern a different construction.

Validation

python3.12 -I -B scripts/research_package.py --package research/qq16-native-dirty-fault

Expected PASS_QQ16_NATIVE_DIRTY_FAULT, eight passing behavioral groups, and PASS_REGISTERED_RESEARCH_PACKAGE.

• Exact identities for all 2,880 used child recipes, all 224 data adapters, all core scalar frame accesses and 1,440 zero residual-boundary values.
• A separate literal execution path covers 26,099 events and 364,544 endpoint amplitudes without frame reduction. It shares the primitive vector evaluator; this is not independent human review.
• Rejecting phase, adapter, normalization, source, fault-sign and fault-location controls; the earlier fault damages 102 helpers. Exact/absolute-error and zero-dirty controls preserve legitimate escape routes.
• Registered CPython 3.12.14 replay, optimized CPython 3.14.7 controls, read-only Git-free export, selected baseline bindings, metadata and all 16 committed package/input byte checks pass locally.

Eight new files, approximately 1,300 lines; no vendored predecessor package, frozen-input edits or shared verifier changes. Full selected cold replay, hosted CI and formal-kernel review are not claimed. The tensor argument is a written consequence of exact vector identities, not a finite-depth experiment.

Credit

Arham Tahir/PR2995 supplied the scalar fault; Nick Petalas/PR2965 supplied the producer-error contract. Credit Ali Romman and the QQ16 team, Sagar Thakrar's native Clifford semantics, DLamdata's selected native-frame lift, ywz65, starship-s, Ross Richardson, Douglas Colkitt and the retained supplier lineage. The general common-frame conjugation lemma is predecessor work. Prepared by Tekalign (@tekuuu) with substantial OpenAI Codex assistance. Apache-2.0; no independent human review claimed.

Open next questions

Pin a finite F0 numerical producer, prove exactness where its grid excludes this fault, and bound the remaining native-child errors using its actual dirty ranges and complete physical bill. Grid/overflow, physical hosting/storage, the paid F17 consumer and supplier admission remain open. The package README gives the precise proof and handoff; no further construction expansion is proposed.