Verify shared-donor kernel improvements on pinned weighted and reordered words

[maximussthegreat]

Additive fresh PR325 certificate: 480 partner-mix retimings

The new sibling research/gen5b-shared-kernels/p10-pr325-partner-retimings/ derives exactly 480 partner setup retimings, from rank 2 to rank 18, directly from immutable PR325 head 0eca9340a3df6141b8e71a41638c3937b3522888. The resulting conditional saving parameter κ is 0.000770612425424136, improving the published paid892-presink value by 0.000000604335525339.

Credit: the replacement word and four new PR325 producer/data changes belong to DreamingOfClouds, with declared Anthropic Claude assistance. The acquired, hash-pinned original NOTICE preserves the detailed earlier contributions and assistance disclosures. The general retiming/descent mechanism remains credited to Rohan Arun's PR287 work and Anthropic Claude assistance. Our contribution is the fresh derivation of these 480 choices, independent complete source/local/global verification, fresh bank allocation, exact invoice and price, and portable packaging, with substantial OpenAI assistance. No old role/frame/chart or retiming-selection data is imported. Upstream programs are retained only as inert text.

The arbitrary-dirty argument explicitly separates address geometry from payload values. Rational address matrices induce permutations of complete F2-valued role arrays; ADD is pointwise XOR. The pinned common-array-frame theorem in proof/primary/03-motifs.tex, lines 146–185, applies after checking every common gate frame, exact edge quotient, fixed endpoint, unchanged logical scalar operator and complete COPY/transform/read/erase episode. These finite premises are independently verified on all 398,660 local records and 12,885 rational nondegenerate frames. The exact source/frame geometry includes all 480 changed setups, 960 source chains, 40,125 MOVE containments and 563,928 regular-source supports.

The global audit reconstructs all 119,604,000 mapped frame records, 300 namespaces, 720 phases, 960 route conjugacies and 100 positive copied-work overrides. Actual projector/conjugacy identities and negative-direction controls pass. All 2,430,000 role/replica/stage assignments fill complete banks, with no completion children. Two independent, equivalent endpoint-chart constructions need at most 139 and 141 factors, both within the unchanged 150-factor allowance.

The literal profile is stock 658,800; 14,514,000 children; rank mass 65,757,600; deficit 122,400; maximum rank 42 < 50. The finite coefficient is 24,857,617,667,871,401. Exact full-fallback pricing, all 47 strict assembly inequalities and rejection of the adjacent κ grid point pass. The portable aggregate has 40 regression tests and 28 compact receipts, with 21 bank and 12 emission negative controls. All 12,884 local basis-mutation controls reject. Thirty-two concrete artifacts match the independently frozen originals byte-for-byte.

The publication baseline is head def71b9eb828aaa31d5e9602b6fc10ac9b844c12, tree 04d58da0bcb5b9c51b31c6f431d820901e51464b. All 354 previous contribution files remain unchanged. Immutable source acquisition and replay work from any working directory. Generated tables and streams are not committed.

All new finite gates are discharged conditionally. Production primitive/compiler realization, uniform eligible-prime and weighted routing interfaces, restored rows, ordinary leaves, complex/full-C, precision/recovery and analytic all-size hypotheses remain inherited. This is not an unconditional all-size theorem or an optimality claim. See the sibling README, proof, exact frame contract, provenance, independent adapter review and deterministic receipts for the precise scope.

---

Additive follow-on: four retimings and three kernels on paid892

The new sibling research/gen5b-shared-kernels/p10-paid-completion892-presink/ independently composes four fixed PR322 retimings and three fixed PR323 rank-two kernels with our admitted 892 matching and changed 60-replica paid packing. The resulting conditional saving parameter κ is 0.000770008089898797, improving the paid892 value by 4.210027284e-9.

Credit: eumemic supplies both the four PR322 selections and three PR323 selections, with declared OpenAI Codex assistance. The original retiming/descent mechanism is attributed to Rohan Arun's PR287 work and Anthropic Claude assistance, as credited by the source. We contribute the independent recomposition, new paid packing, source/global verification and exact price on our matching. Upstream source programs remain inert text.

The finite checks reconstruct 389,222 local records, all 15,041 used frames, all 133 retained anchor events and 4,141 crossings. They verify 589,692 integer source checks, 116,772,600 global frame records, 921,600 boundary records, all 300 namespaces, three changed endpoint charts and 725 phases. Each of the five final banks has 80 real coordinates and a fully paid rank-20 complement. All 100 positive copied-work overrides are included, with actual rational projector/conjugacy tests and negative-direction controls.

The literal profile is stock 658,255; 15,006,605 children; rank mass 65,703,100; deficit 122,400; maximum rank 42 < 50. All fees, the exact finite coefficient 25,478,454,083,580,596, all 47 strict assembly inequalities and adjacent-κ rejection are independently checked. The portable aggregate has 30 regression tests, 17 embedded hard controls and 15 compact new receipt hashes, plus the pinned predecessor's 33 tests and 31 receipts.

The baseline is published head 4c40c0151cf8317e57a3e7759d495fb84653bc74, tree 25d778000cfcc133d7bc93d7e04d7179a4b6e144. All 323 existing contribution files remain unchanged. Source acquisition and replay work from any working directory; no bulk generated tables are committed.

The new finite obligations are discharged conditionally. Production primitive execution, compiler/prime uniformity, prime supply, restored-row, ordinary-leaf, complex/full-C, precision/recovery and analytic all-size interfaces remain explicit inherited premises. This is not an unconditional all-size theorem or an upstream whole-program replay. See the sibling README, proof, provenance, independent adapter review and deterministic receipts for exact scope.

---

Separate paid-completion experiment

A new additive packet at research/gen5b-shared-kernels/p10-paid-completion892/ tests the 892-pair candidate on immutable PR320 source head 1b37957d1520c80b6ea796bf418e52be5109c2d4. It preserves all 283 existing contribution files and leaves the earlier weighted-890 packet intact.

The paid 49+11 completion fills the remaining 60 arbitrary-dirty coordinates of one bank at each of five stages. The two complete cover sweeps occur after the 60 regular helper sweeps and before that stage's boundary. Their ten additional children are fully charged; no extra closure relation, free padding or fractional physical indexing is introduced.

The resulting conditional saving parameter κ is 0.000770003879871513. The literal profile is stock 658,275, calls 15,002,710, rank mass 65,705,100 and deficit 122,400. The maximum rank remains strictly below 50. Independent exact calculations verify the full fallback moment and all 47 displayed assembly inequalities.

The new integration obligation is checked by concrete emission and independent artifact audits. The full-frame companion additionally reconstructs the actual final local frame word, screens all 133 reorder moves, checks all 15,034 used frames and maps 116,772,900 frame-tagged records through all 300 global namespaces. The source-geometry correction explicitly covers all 480 descended rank-18 setups and all 960 final source chains. Earlier supplementary source-chain language is limited to pre-descent geometry.

The finite evidence remains conditional on the inherited generic primitive compiler, uniform admissible-prime supply, complex/full-C supplier, precision/recovery, primitive and leaf constants, and analytic/all-size interfaces. It does not execute the upstream production compiler or assert an unconditional multiplication theorem. Upstream programs are only read as pinned inert text.

Run the portable aggregate from any working directory:

P10P_INPUTS=/path/to/source-cache \
P10P_OUTPUT=/path/to/fresh-output \
python /path/to/p10-paid-completion892/run_checks.py --check --tests

The portable replay passes 33 regression tests and freezes 31 deterministic receipt hashes. Source acquisition, exact authored dependency pins, compact matching and packing witnesses, and proof notes are included. Expanded tables and binary streams are regenerated in the output directory rather than committed.

Prepared with substantial OpenAI assistance. Existing contributor credits and Apache-2.0 notices are preserved; the new packet's NOTICE.md identifies inherited ideas and the source chain.

Updated p10 baseline: weighted 890 reuse with all 1047 kernels retained

This adds research/gen5b-shared-kernels/p10-updated1047/ and preserves all 249 earlier files in the shared-kernel contribution. The earlier p10 experiment remains unchanged and separately pinned; the newer baseline is now independently reproduced here.

At immutable PR320 head 1b37957d1520c80b6ea796bf418e52be5109c2d4, the conditional saving parameter κ increases from 0.000769553898621543 to 0.000769997773182046, about 0.05767946%. The new matching keeps 890 pairs, changes 63 exact pairs and retains all 1047 installed kernels. No old fixed 19-relation closure is added.

• Stock stays 658,290; banked calls decrease to 250,045 at unchanged rank mass 1,095,110.
• Independent reconstruction checks all 18,648 eligible edges. Rank-prefix matching/cover certificates prove the maximum primary donor benefit at cardinality 890 on this fixed graph. Residual potentials prove the maximum 827 retained original exact pairs within that rank profile. This does not optimize κ globally or establish uniqueness.
• Fresh checks cover all 8,230 old/new alias paths, 1,047 kernel relations and cuts, 120 target relations and 480 closing-frame members, with physical IDs explicitly rebound.
• A complete fresh pre-reorder scalar/COPY replay passes in both directions. All 133 inherited reorder moves have matching semantic anchor and noncommuting-event traces.
• Fresh charts and 2,466,900 explicit role/replica/stage bank assignments pass. The payload is below 2^98; the finite ring/recipe coefficient is 25,476,360,501,102,401, below the retained 2^80 budget.
• Two independent exact rational moment engines certify the bit-root bracket. All 47 retained assembly inequalities pass.

The standard-library-only portable aggregate passed 13 stages, 50 regression/mutation tests and 14 deterministic receipt checks. The 34-file addon contains authored code, a small exact matching delta, a 39-blob immutable input manifest and compact expected receipts. Large derived objects regenerate locally; upstream Python is read only as inert text. Reproduction is documented in the addon README.

Evidence boundary: the original 133 raw record-number crossing proofs and retained transform/compiler, prime, complex-supplier, full-C and all-size interfaces remain explicit inherited hypotheses. The packet does not regenerate the complete final raw frame/record/hash pipeline, instantiate a numerical full-C/all-size cutoff or prove an unconditional multiplication theorem. No global-optimality or current-best claim is made.

The addon NOTICE preserves the PR275 weighted-matching attribution and the PR315/319/320 contributor chain. New material is Apache-2.0 and prepared with substantial OpenAI assistance.

---

Additive p10 update: weighted reuse and 19 shared-donor relations

This adds research/gen5b-shared-kernels/p10/ while preserving all 200 previously published files.

Starting from PR320 at immutable head e983fea0896b2a1d7355983674db7e9bb33a9e32, the conditional saving parameter κ increases from 0.000769198971896986 to 0.000769661484777339, approximately +0.060129%. An explicit 892-pair weighted matching and 19 compatible shared-donor relations are priced jointly.

• 8,221 helper roles; stock 658,905; banked calls 249,085; rank mass 1,096,135.
• All 775 inherited kernel relations/cuts checked, including 66 affected entries.
• Fresh base-word and complete pre-reorder scalar/COPY replay on all 10,148 pre-compaction columns, with seven inert sink spectators.
• All 120 target dependencies and 480 member closing-frame paths checked.
• All 133 reorder transports checked using complete semantic anchor/noncommuting traces.
• 47 endpoint/quotient and 57 seam chart programs checked exactly.
• All 2,466,300 role/replica/stage bank assignments checked.
• Fresh F = 33,129 and B = 984,792; the 91-bit payload fits the retained 104-bit cap.
• Finite coefficient 25,415,456,130,362,201, charging all 116 added unit ADDs without credit for removed reads.
• Exact bootstrap, all 47 assembly inequalities and adjacent-grid rejection pass.

The unchanged-matching 15-relation case remains reproducible as a fallback. The 49-file addon uses a single 40-blob immutable input manifest; upstream Python remains inert. Large derived tables regenerate locally.

The portable aggregate checks 22 stages and 24 deterministic receipts. There are 44 focused tests, including source-corruption, matching, geometry, scalar, bank and arithmetic controls. Reproduction from an unrelated working directory is documented in p10/README.md.

Evidence boundary: the original 133 raw crossing admissions and retained compiler, prime, complex-supplier, full-C and all-size interfaces remain explicit inherited hypotheses. The new work does not regenerate the complete final raw-frame/hash pipeline or prove an unconditional multiplication theorem. It makes no global-optimality or current-best claim.

At the 17:38 UTC publication check on October 10, PR320 had advanced to 1b37957d1520c80b6ea796bf418e52be5109c2d4 and its immutable pins advertised κ = 0.000769553898621543. That newer result has not been independently reproduced here; this packet remains pinned to the earlier baseline above.

Attribution to the PR275 weighted-matching method and PR315/319/320 source chain is preserved in p10/NOTICE.md. New material is Apache-2.0 and prepared with substantial OpenAI assistance.

---

Added: shared156 transport through pinned PR306 reordering

This additive weighted306/ packet reproduces PR306's baseline arithmetic at immutable head 0314371983b8837f01723f5af2c70217f75b01cc and checks composition with the already published shared156 witness. All 174 earlier files remain byte-identical.

• Pinned PR306 baseline: κ 0.000751458925311258
• Combined conditional result: κ 0.000751478472895775, about +0.00260128%
• Stock 1,229,275; priced calls 483,015; rank mass 2,454,150
• Same 156 relations, 197 donors and 353 members; 18 legal later rank-22 visits, unchanged initial quotients and bank endpoints
• All 281 exact chart inverse programs and 706 role demands rebound; full prior five-stage role-bank certificate hash-bound
• F ≤ 438151, B ≤ 14104135; payload 110 bits: legacy 104 fails, explicit 112 passes
• Displayed finite coefficient 90417469526202901 (<2^80), three exact bootstrap steps and all 47 displayed assembly inequalities pass

Verification boundary: PR306's raw crossed-interval and complete-word admission are inherited. This packet does not regenerate its complete raw word or independently replay all 239 original crossing checks. It proves that every added or deleted shared156 scalar gate lies outside those full intervals, checks all changed selected-role paths, and transports exact signed/unsigned commutation. The earliest interval starts at 690991, after the conservative prefix bound 656881. The unchanged compiler, physical decoder, prime/recovery, complex and all-size interfaces remain conditional; the displayed finite invoice does not instantiate the unknown full primitive constant. No current-best or optimality claim is made.

Reproduction is documented in weighted306/README.md: verify306.py --check --prior-inputs DIR305 --inputs DIR306 --output FRESH_DIR. The addon pins 13 new inert inputs by immutable URL, Git blob, SHA256 and byte count, and reuses the existing 39-input manifest and published156 evidence. No upstream producer or other upstream program was executed.

Validation: 36 focused tests passed in 71.650 seconds, including source/witness corruption controls, missing-visit rejection, exact signed/unsigned and COPY commutation checks, and finite/chart mutations. Four post-freeze controls also passed. The independent final portable aggregate passed in 67.355 seconds using 13 newly acquired inputs and 39 clean prior inputs, matching all five compact receipts. Earlier70/166 suites were not rerun because their files are unchanged.

Prepared with substantial OpenAI assistance, Apache-2.0. PR306 reorder work is credited to Chafik Boukhalfa/Anthropic Claude; the retained PR305, George Lydakis's PR275, Rohan Arun's PR300, PR299/296 and earlier contributor/assistance chain is preserved in weighted306/NOTICE.md and all existing notices.

---

The earlier fixed-baseline results and their original verification limits are retained below.

Additive PR305 port: two fixed, independently checked cases

Adds research/gen5b-shared-kernels/weighted305/. All 136 earlier files, including the complete weighted275 and original PR302 evidence, remain byte-for-byte unchanged.

Pinned PR305 (a17c42903bbe8d9d26c4bc712d7216a7176237e9) independently reproduces κ = 0.000751140529238897. The complete 154-group port gives 0.000751160050810041. Exactly one restricted third-line reclosure gives the 156-pivot case, 0.000751160060489770, approximately +0.00260021% over that pinned baseline. These are historical comparisons, not a current-best, exclusive-priority or global-optimality claim. At publication, newer PR #306 (head 0314371983b8837f01723f5af2c70217f75b01cc) claims κ = 0.000751458925311258, stronger than this update's 0.000751160060489770; that newer claim has not been independently reproduced here.

The bounded change preserves the prior 136-pivot line16/17 and 18-pivot line6/7 groups, then recloses only the five original line2/3 relations surviving six newly occupied roles. It retains three and applies the declared parity trim, leaving two. No additional lines or bases were explored. Compact subset indices regenerate the exact SHA-bound witnesses from the already published 166 witness.

Fresh checks use the actual eight sinks and 16 redirects, rather than importing the previous seven-sink norm. For 156:

• All 353 selected paths, including 102 reused donor slots, and 2,468 exact containments pass; the conditional composition identity is checked in both directions on all 18,952 pre-compaction physical columns.
• 281 exact chart programs bind 706 role uses; retained factor/normalizer bounds pass.
• All 15,424 helpers in 60 replicas are explicitly assigned: 925,440 bank addresses per stage, 4,627,200 total.
• Stock 1,229,275; calls 483,015; rank mass 2,454,150; all 47 rational inequalities and next-grid rejection pass.
• Fresh F=438151 and B=14104135 give a 110-bit payload. Legacy104 fails explicitly; the separately reviewed fixed112 certificate passes.
• The displayed finite coefficient is 90,417,469,526,202,901 (<2^80), charging all 2,545 setup and 2,545 inverse ADDs without removed-read credit.

The final independent clean run passed all 47 focused tests in 122.774 seconds, including the auxiliary-manifest regression and eighth-sink mutation controls. The complete fresh-input aggregate passed in 109.188 seconds, with all 39 immutable inputs independently downloaded and verified. Both full case replays match their frozen receipts. The changed-obligation checks remain conditional on the inherited physical decoder, source-span/chronological admission, compiler, primitive constants and all-size interfaces. No literal whole-upstream regeneration, concrete value of unspecified C_full, or full multiplication theorem is claimed.

Reproduce with Python3.11+ standard library, assertions enabled:

cd research/gen5b-shared-kernels/weighted305
python3 -B acquire_inputs.py --output /tmp/weighted305-inputs
export PR305_INPUTS=/tmp/weighted305-inputs
export PR305_OUTPUTS=/tmp/weighted305-results
python3 -B -m unittest discover -s . -p 'test_*.py' -v
python3 -B run_checks.py --check --inputs "$PR305_INPUTS" --output "$PR305_OUTPUTS"

All 39 source inputs are pinned by immutable URL, byte count, SHA-256 and Git blob hash. Upstream Python remains inert .py.txt; no upstream program or repository-wide make verify was executed. Parent exact arithmetic and sibling pure chart/closure/acquisition helpers are reused. Full generated charts and bank tables are omitted from the diff.

Attribution remains with the PR305 composition, George Lydakis/PR275 weighted matching and predecessor integration, Rohan Arun/PR300 restart packing and descent, Chafik Boukhalfa/PR299–296 and the retained source chain. The new NOTICE links immutable notices and preserves source-specific licensing. New material is Apache-2.0 with substantial OpenAI assistance; separate AI-assisted checks are not human peer review or an endorsement.

---

The earlier pinned results and reproduction details below remain historical evidence.

Additive weighted-word update

Adds research/gen5b-shared-kernels/weighted275/ to this same research contribution. The original 91-file PR302 evidence and manifests remain byte-for-byte unchanged.

The new fixed case uses PR275 at 10b40041d4ab8a6610083e95bc571aee3468bca2, plus 166 shared-donor pivots. Its conditional κ is 0.000751097723288767, versus the independently reproduced pinned baseline 0.000751075693834608 (about +0.00293305%). A separately verified 70-pivot port remains as fallback. These are historical pinned comparisons; newer pending work may have a stronger headline. No current-best, priority or global-optimality claim is made.

The corrected search admits unused physical reuse-donor slots while excluding recipient keys and occupied/descended/early-restored streams. The declared experiment covers 276 contrast lines and exactly two overlap-priority orders with shared-donor closure and deterministic parity trimming. The final witness has 166 pivots, 210 donors and 376 affected physical slots, including 105 reuse donors. The original final12 endpoint saving does not transfer to this word and is explicitly excluded.

Reproduced changed obligations

• All 18,952 physical columns in both directions for the conditional composition identity; all 376 aliased paths and 2,561 exact containments.
• 304 exact rational chart programs bound to 752 role uses; retained chart and combined normalizer bounds remain valid.
• All 15,425 helper roles in 60 replicas, 925,500 concrete bank addresses per stage and five disjoint stages.
• Stock 1,229,320; 483,150 calls; rank mass 2,454,240; all 47 exact assembly inequalities and adjacent-grid rejection.
• Fresh sink-aware F=433193 and B=14098095. The payload bound is 110 bits: the original 104-bit assertion fails. A separate, explicit 112-bit fixed-recipe certificate passes, with a static audit of the pinned source and retained bit/compiler quantifiers. No hidden relaxation of the original checker is claimed.
• Displayed finite coefficient 90,433,795,441,918,801 (<2^80), charging all 2,587 setup and 2,587 inverse ADDs without removed-read credit.

The inherited physical decoder/source-span admission, unchanged compiler/primitive interfaces and all-size framework remain hypotheses. The displayed bill does not instantiate the unspecified full primitive constant. Full-column composition checks are not literal regeneration of the complete upstream scalar stream, and finite success is not formal verification of the multiplication theorem.

Reproduction and provenance

Python 3.11+ standard library only, with assertions enabled:

cd research/gen5b-shared-kernels/weighted275
python3 -B fetch_inputs.py --output /tmp/weighted275-inputs
export PR275_INPUTS=/tmp/weighted275-inputs
export PR275_OUTPUTS=/tmp/weighted275-results
python3 -B -m unittest discover -s . -p 'test_*.py' -v
python3 -B run_checks.py --check --inputs "$PR275_INPUTS" --output "$PR275_OUTPUTS"

All 39 immutable inputs were acquired into a fresh directory and verified by byte count, SHA-256 and Git blob hash. No upstream program is imported or executed; Python source is stored as inert .py.txt. Large downloaded inputs and generated full charts/bank tables are omitted from the diff. The original PR302 reproduction commands remain unchanged.

An independent clean verification run passed all 44 focused tests in 102.155 seconds, including independent mapping, relation, alias chronology, chart, bank, replica, scalar-bound and old-cap negative controls. The full two-case aggregate and discovery replay passed in 153.906 seconds. A separate fresh acquisition independently verified all 39 immutable input files. Repository-wide make verify and upstream producer programs were deliberately not executed; only the newly authored checks ran.

Credit is retained for George Lydakis and the PR275 weighted-matching contribution (including Ody's matching research), Rohan Arun's inherited PR300 descent, Chafik Boukhalfa's PR299 package and the full source chain. New material is Apache-2.0 and was prepared with substantial OpenAI assistance. The new NOTICE links the immutable original notices; independent AI-assisted checks are not human peer review or an endorsement.

At publication, PR #305 claims conditional κ = 0.000751140529238897, stronger than this update's 0.000751097723288767. That newer claim has not been independently reproduced here.

---

Original pinned PR302 contribution

Summary

Add a standalone verifier and construction note for 70 shared-donor rank-one kernel pivots, composed with pinned PR300 and the 12-helper co-retiming from PR301. A fully checked 72-pivot fallback and PR299-based arithmetic checkpoint are retained separately.

Under the retained source/compiler/all-size interfaces, the combined 70 case gives:

• Exact conditional kappa: 749940157923873/10^18 = 0.000749940157923873.
• Bit coarse saving: 750502990558158/10^18.
• Literal stock 1,229,560; unreplicated calls 483,425; rank mass 2,454,720; deficit 4,400.
• 70 pivots, 87 distinct donors, 157 affected helpers, and 601 setup/restore shear pairs.
• Displayed finite coefficient 90466979917115701 < 2^80.
• Corrected sink-aware payload 12927738210344389271740351838400 < 2^104, with strict positive gap 7354671393307281152206899447616.

The separate 72-pivot fallback on PR300 plus co-retiming gives kappa 0.000749936120273400, stock 1,229,555, calls 483,455, rank mass 2,454,710 and a 102-bit payload. The PR299-based 72 checkpoint gives 0.000749917402998347. No global priority or optimality claim is made.

Comparisons here use the pinned older heads; newer upstream revisions exist, and this PR does not claim the best current bound.

What is new

Pivots may share donors on the same rank-one entrance line. Pivot and donor sets remain disjoint, so the shear incidence matrix squares to zero and the setup/restore factors commute. Exact response relations, a reproduced common read cut and complete local frame paths certify the changed bit construction compositionally under the inherited decoder.

The packet includes independent role-to-bank allocation, not just inventory arithmetic: all 15,427 helper roles and all 60 replicas are assigned, totaling 925,620 assignments per stage and 4,628,100 across five disjoint stage namespaces. The 70 case uses 161,432 full width-120 banks per stage. Its 171 exact chart programs cover all 314 new role uses plus the 18 co-retiming endpoint programs, with at most 489 factors, within the retained 576/815 allowances.

PR300's 89 scalar events match original forward/old-kernel incidences; their 154 helper operands are disjoint from the broader 84-stream co-retiming closure and all new-kernel members. All 440 restoration and seven sink entries match PR299 semantically. Exact determinants additionally verify 65 distinct selected PR300 bases. The finite bill includes all 1,202 new scalar additions per stage and the actual sink redirects/mixed target paths.

Immutable dependencies and diff scope

The change adds only research/gen5b-shared-kernels/ against main:

• PR299 head 5f6bd3fbc0e6796dd31263cef1f06dc968186f64, package research/five-stage-gen5b-kernels-restorations-sinks.
• PR300 head fd516176fd068e4fa7af14b40cd878ce8c305296, package research/five-stage-gen5b-full in rohanarun/integer-mult-bounds.
• The included co-retiming component is PR301 head ebd6a08ed7d517387794c066490aa02aadcc78d4, included locally for standalone reproduction.

Pending PR data is explicitly fetched using 16 immutable byte-count/SHA-256/Git-blob pins. A main checkout alone is not assumed to contain these inputs. This does not copy those pending PRs' upstream source bundles into the main diff, alter PR301, change PR294, or change main's selected certificate, production guards, upstream/ or manuscript patch.

Only independently authored Python runs. The one fetched upstream .py file is read as inert loader-order text, never imported or executed. Large raw inputs and full generated scalar/chart/address tables are excluded from the committed files.

Reproduction

Python 3.11+ standard library, with assertions enabled:

cd research/gen5b-shared-kernels
python3 -B fetch_inputs.py --destination /tmp/shared-kernel-inputs
export CORETIME_SOURCE_DIR=/tmp/shared-kernel-inputs
python3 -B -m unittest discover -s . -p 'test_*.py' -v
python3 -B run_checks.py --check

run_checks.py --output-dir /tmp/shared-kernel-results --check also generates complete factor programs and deterministic bank address tables; add --write-streams for the full co-retiming suffixes.

The aggregate checks include all 18,954 columns of the compositional bit identities; the full 55,985-event co-retiming suffix over F2/Z/absolute majorants in both directions; changed chronological frames; exact chart factors; complete role assignments; conservative finite/scalar bounds; exact moment brackets; three bootstrap steps; all 47 strict outer constraints and adjacent-grid rejection. Negative controls cover wrong response relations, missing setup/restoration, frame overlap, malformed arithmetic, omitted sink edges, changed source pins, chart/bank drift, duplicate roles and equality at the strict payload cap.

The final clean run passed all 135 distinct targeted tests on Python 3.12.14 in 280.020 seconds. All 27 committed JSON receipts were independently reproduced exactly as data by the portable aggregate from /tmp. Patch application and reconstructed-file hashes are also checked. Only independently authored verification code was executed.

Evidence boundary

No additional open changed-stage obligation is identified by these checks. The original PR299 decoder, prefix/source admission and original row norms remain hypotheses. PR300's complete chronological/source-span admission is inherited: matching events, support and basis determinants do not reprove it. Unchanged chart/normalizer, cover, routing, prime, recovery, setup, finite-compiler, complex-supplier and analytic/all-size interfaces remain inherited too.

The shared-kernel all-column proof is compositional under the old decoder, not a literal replay of the entire upstream physical word. The bit interface stays F2; the stronger integer identity belongs to the co-retiming suffix only. The displayed finite coefficient does not set unspecified constants in C_full to zero. Upstream make verify, make entrance-bank-verify, complete physical admission and the full all-size theorem were not run. This is conditional research evidence, not formal verification of the multiplication theorem.

Attribution and assistance

This builds on Chafik Boukhalfa's PR299 package and Rohan Arun's PR300 descent, preserving their immutable source notices and inherited lineage. Credits include DreamingOfClouds, Rohan Arun, eumemic, utcorvusvolat-dotcom, Dugongue, Evan McKinney, Henry Grant, Jacob Sussman, ikeboy, icekylinx, James Chang, sennemmi and their credited dependencies. PR299's substantial Anthropic Claude assistance disclosure is retained. The prior PR301 co-retiming is explicitly identified rather than presented as a new mechanism here.

New material is Apache-2.0 and was prepared with substantial OpenAI assistance. Separate AI-assisted checks are not human peer review, an OpenAI endorsement or proof-assistant verification. See NOTICE.md for the full attribution boundary.