Skip to content

Conditional κ = 0.001673619538656538 from a verified two-role T16 swap - #3338

Merged
Mister-Meeseeks merged 5 commits into
CrocSwap:mainfrom
petalas:research/20261017-batch5-worker-1-kappa
Oct 18, 2026
Merged

Mister-Meeseeks merged 5 commits into
CrocSwap:mainfrom
petalas:research/20261017-batch5-worker-1-kappa

Conversation

@petalas

@petalas petalas commented Oct 17, 2026 •

Copy link
Copy Markdown
Contributor

Summary

 unchanged selected PR1094 word, suppliers, chronology and full physical bill
- idle ranks (2,1), PR3308 κ = 0.001673618490895016
+ idle ranks (3,0),          κ = 0.001673619538656538

Exchange two whole helpers between capacity-16/32 bins. Strict concavity proves the moment gain. The complete bounded census checks 19,372 different-capacity pairs and finds eight legal choices. The committed witness relocates two roles and changes four additional offsets.

Exact gain is 523880761/500000000000000000 over PR3308 and 47967078981/50000000000000000 over selected κ 0.001672660197076918. This is a strict gain at the unchanged selected conditional finite-composition scope, not a global-best or complete multiplication theorem. Selected-result.json is unchanged. PR3308's archive-only merge is not scientific admission.

research/t16-two-role-moment-swap/ contains the map, construction/proof, exact reproduction, failure controls, source closure and three immutable independent supporting bundles. PR3308 is reconstructed from canonical data plus its original delta; cold replay does not need an unrelated cached Git object.

Evidence

  • Before: authenticated PR3308 map and complete selected paid invoice. After: unchanged eleven-stage strict admission, controls 69/12/6/15, 69,120 roles, 144 calls and actual 10,008 charts pass. Every invoice field matches except three freshly rebound source hashes.
  • Registered replay runs the final immutable v3 code-disjoint scorer on fresh strict outputs, then the final invoice reviewer against that fresh score and exact copied source. Reports match their explicitly source-rebound expected bytes only after whole typed invoice/assembly equality and exact live hash-chain checks. Eleven normal/optimized arithmetic groups, six construction controls, thirteen invoice/source controls, twelve normal/optimized construction/diagnostic groups and nine normal/optimized transport groups pass.
  • Both supplier caps, native/conservative bootstraps and all 47+47 strict outer inequalities pass. Complex still binds. The fixed-word cap is 1690054109089/1001690054109089; this witness realizes about 99.1949333% of it, without an optimality claim.
  • python3.12 -I -Werror -B scripts/research_package.py --package research/t16-two-role-moment-swap, committed-byte preflight and the final 117-file Git-free committed export pass locally. All exported bytes remain unchanged.
  • Hosted run39075974678 completed SUCCESS on exact final head f2e33dd2c07caf5dac2ee42505852a081d8fa7ca: both plan and full registered replay pass. Actual hosted witness bytes and every fresh independent-gate marker were checked.

The first hosted replay exposed ordering differences in the ten JSON artifact-map entries, not a changed invoice value. Actual invoice, assembly, owner and controller bytes were preserved and independently inspected before the repair. The wrapper now proves complete typed JSON equality with no cost/key exclusions, rebinds only the explicit raw dependency hashes, and rejects changed costs, caps, inequalities, unrelated hashes and bad bindings. Immutable v3 sources/reports and raw strict outputs are unchanged.

Canonical refresh 6f100d4bb5083f8c35d96fd95a539811c62f245c leaves all 43 scientific input blobs, priority revision 6 and the selected result unchanged. Frozen scientific base remains 463113d675af7410ac91d53a4fed51ec6494eb9a.

Native ancestry is source-bound reuse, not a fresh full cold rebuild. All inherited all-size/compiler, common-label/mutable-source, whole-call hosting, precision/tape, exact recovery, analytic and bit H1-H5 premises remain. No independent human or formal-kernel review is claimed.

Credit Gonzalo Mier/PR3308, Philippe Van Velzen/PR2169, Rohan Arun, Jeremy Kothe, Hao Yan, Josie Elliston, Sagar Thakrar and the retained compiler/native/bit/moment lineage. Exact notices and supporting source manifests are preserved. Open next questions are in the package.

Merge Danger

Door: two-way. This adds one isolated research package and does not select a frontier or modify shared scientific code.

Blast radius: package. One immutable scorer retains a documented nonbehavioral Ruff F401 warning; its checked source bytes are preserved rather than silently edited.

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