Skip to content

Bind the QQ16 F0 rounding witness to physical native execution - #3067

Merged
Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
RossRichardson:research/parallel10-20261016T212207Z-06
Oct 16, 2026
Merged

Mister-Meeseeks merged 1 commit into
CrocSwap:mainfrom
RossRichardson:research/parallel10-20261016T212207Z-06

Conversation

@RossRichardson

Copy link
Copy Markdown

The concrete QQ16 F0 scalar-rounding failure in #3036 survives the physical native frame and residual adapters under an explicit mixed arithmetic model: scalar ADD components round nearest-even, while completed native operators remain exact. At significand precisions 8, 24, 53 and 113, the literal 26,099-event native word has exactly two rounding sites and five wrong data fields, although all 480 physical helpers return exactly and all eight private fields return zero. The largest Hilbert error equals the original data norm.

This closes #3036's stated physical binding gap for these four finite cases. It does not round native-child interiors, establish an all-input producer guarantee, supply a complete recursive invoice, admit a supplier or change kappa. The original scalar witness is credited to Arham Tahir/#3036; the unchanged exact vector evaluator is retained from Tekalign/#3017. Nick Petalas/#2965 and the Ali Romman/QQ16/native lineage retain their attribution and premises.

Adds only research/qq16-f0-physical-rounding/: source pins, exact receipt, literal physical execution and integer nearest-even rounding, the written frame/norm argument, and eight behavioral groups. Internal independent mathematical review checked the marked frames, signed continuation responses, zero residual boundaries, frame rounding commutation and scope; it shares the retained primitive evaluator and is not independent human or formal-kernel review.

Validation:

  • Registered CPython 3.12 replay passes, with four complete physical executions and eight optimized behavioral groups.
  • Affected final tests pass after adding typed receipt corruption checks; optimized CPython 3.14.6 passes a Git-free export of the final verifier and immutable receipt.
  • The independent nearest-number oracle covers 2,575 decisions; clean and wider-same-input controls, separate rounding sites, dyadic sign/scale, native primitive and occupied nonconstant adapter controls pass.
  • Metadata/source pins, warning-as-error compilation, selected baseline bindings and whitespace checks pass. The selected cold replay and hosted CI success are not claimed; no CI rerun or dispatch is requested.

Reproduce:

python3.12 -Werror -I -B scripts/research_package.py --package research/qq16-f0-physical-rounding

Scientific and submission main: d8e0f83f00191df62174b641e9fff055c357f641, priority revision 6, refreshed at 2026-10-16 21:34:37 UTC. Pre-submission overlap fully paginated 304 open PRs including drafts, 1,817 Discussion records and 2,440 closed PR bodies; source heads #3036/#3017 are unchanged. Metadata snapshots are non-atomic and exclude arbitrary diffs and unpublished work. Canonical investigation.

The next useful task is an actual finite native producer's grid/dirty-range/internal-rounding and complete paid physical contract. Exact-grid arithmetic remains a valid escape. Family decision: narrow and stop. New work by Ross Richardson with substantial OpenAI Codex assistance, including an internal rounded-networks specialist; Apache-2.0, predecessor notices preserved.

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.

2 participants