KB1c: admission-ready evidence for the typed-window z4 typed-echo schedule (no new kappa)

[navinor]

KB1c: admission-ready evidence for the typed-window z4 typed-echo schedule (no new kappa)

Agent: Aaron Wang, with help from opencode (substantial AI assistance; no independent human review claimed). No model is named. Base main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2 (assignment-pinned; canonical main moved to 6bd33bc0b87a01e9c9d8b1c01576f3ddbcaee850 during the session — every pinned chain input is byte-identical there except the task catalogue, which this package pins to the base revision; see Baseline below); priority revision 4; selected record unchanged; no new kappa and no supplier admission.

Package: research/z4-typed-window-admission-kb1c/. Claim: #1166 (comment)

What this is

#2464 built the first hosted supplier instantiation of the certified z4 typed echo (whole-construction pricing, a < D/E) and stopped on one hosting obstruction. #2595 removed that obstruction: legal per-call sequential immutable COPY/ERASE windows with per-call totals delivery keep a positive whole-construction saving at k=2 (384 records literal / 304 typed against the priced 608 baseline, typed relabeling priced at zero extra). This package carries the schedule the last mile: it runs the pinned PR 1087 chain's own admission/consistency machinery against the typed-window declarations as far as non-invasive execution reaches, produces the exact evidence bundle (inputs hashed, boundary contracts, dirty-column counts, the priced ledger), records the exact stage and error of every hard blocker, and states the named maintainer decisions that remain. A maintainer can act on this in one step; nothing in the frozen chain, the selected record or the selected frontier is changed here.

The frozen chain runs fresh, unmodified

The full strict T16 replay (research/gauge44-slots-t16/verify_t16.py) runs end to end on the retained local word of the pinned PR 1042 replay archive: all eleven stages (restore-parent, bind8, lift16, bind, emit, audit, invoice, assembly, admission-controls, bit-supplier-controls, independent-t16), verdict PASS_GAUGE44_T16_STRICT_GLOBAL_REPLAY at kappa 1672188119205421/1000000000000000000 (0.001672188119205421 — the PR 1042 composition, not the selected 0.001672660197076918). Controls run fresh: 69 admission controls (every mutation rejected, none accepted), 12 bit-supplier controls (derived deficit 10272), 15 controller mutations, 6 independent-t16 controls, 47 strict outer margins all positive. Binding inventory: 69,120 roles, 144 complete calls, prefix 397 at charge 512, physical row coefficient 16,469 below 20,161. The owner receipt reproduces the archived receipt byte-exactly (SHA-256 888d63e292eed39215fe9f1440b15018f71507e13132f0e02bf0c636d56908e6); the invoice receipt and stage outputs are semantically identical to the archive (byte differences are one receipt's JSON key order and the absolute paths independent-t16 records).

Evidence bundle

Inputs hashed (RESULT.json): all chain code/data/engine files, the PR 386 contract text, the frozen replay archive and its eight receipt members, the task catalogue and selected record, and the vendored PR 2595 machine receipt (inputs/PR2595-RESULT.json, SHA-256 a740acb4727ca5141154617e47986de41a3c7cfbfdac2ec2c5885eace55552e7, PR head 533bb00517ad0e9206ec34d2026a3977a2dc9a16, fetched byte-exactly).

Boundary contracts (each located at its exact enforcing line): the PR 386 sentence "COPY/ERASE windows remain sequential and immutable" (research/split-centres-global/proofs/PR386-BIT-GEOMETRY.txt line 54); wrapper counts bound to entering helpers (code/assemble.py line 102); wrapper factors charged per call (line 126); the selected deficit literal owner deficit == 5120 (line 188); the supplier profile pins expected.rank_mass (line 67), expected_profile (line 85), complete_contract (line 92), expected_local (line 97), expected.global_profile (line 182). The provider local receipt has 13 fields; none names a typed relabeling, a window width or a per-pair totals sink.

Dirty-column counts: chain receipt 704 = 2x112 + 480 all-exact columns per call, 480 dirty helper columns, all restored true, 69,120 = 144 x 480 hosted roles; typed-window schedule at k=2 covers all 56 basis columns (every dirty state), with 0 dirty rows at block-2 entry for the legal per-call schedules (24 for the fused word, 16/24 for row-minimal deferral).

Priced ledger (width m = 73, live stock W = 14,336; exact moment acceptance is the chain's own engines/moment.py + engines/base_two_moment.py predicate, accept at the endpoint and reject at endpoint + 1e-24):

ledger	D	endpoint (chain oracle)	cross-check
chain receipt baseline	5120	418747253984418082117/250000000000000000000000	the chain's recorded complex_lower, accept/reject reproduced
published baseline	5120	67018507072324140917/40000000000000000000000	equals the published root 0.0016754626768081035
chain shared (widened window, n_8 576)	9728	796085643023976876837/250000000000000000000000	above the certify ceiling (obstruction 1)
published shared (widened window)	9728	1592621207654949794599/500000000000000000000000	equals the PR 2595 enclosure [0.003185242415, 0.003185242416]

The a < D/E screens remain screens: the chain engine's moment lower bound at the pinned D_over_E_lo already exceeds 1 (control D_over_E_screen_not_acceptance). Scalar side (PR 2595 machine receipt): k=2 pair 384 literal / 304 typed against 608 (224/304 primitives saved). One bounded pivot (k=5+5): phase-indexed readers 1,456 separate / 1,408 joint. The fixed-reader family is qualified by #2626 (1,616 separate / 1,672 joint, 56 extra), so the widening decision should be priced on the phase-indexed family.

Scoped obstructions (exact stage and error)

1. Moment certification bracket. The widened-window ledger's root (about 0.0031852) lies above the engine's hard-coded search ceiling hi = Decimal('.003') (engines/moment.py line 78); the final bracket assertion at line 86 fails (AssertionError: assert lv[1]<1<hv[0]). The engine's moment() acceptance oracle itself accepts the ledger at its own grid endpoint (found by bracket bisect on that oracle).
2. Owner deficit literal. With the widened ledger declared consistently (owner profile, rank mass, deficit 9728, per-call rank-8 children 8 to 4, adapter charges, pervertex and finite envelopes re-derived) and every SPEC pin relaxed, code/assemble.py still hard-codes owner['deficit'] == 5120 (line 188, empty AssertionError).

Intermediary rejections, all recorded with their exact pins: the owner-ledger-only declaration dies at the per-call profile/idle consistency (line 130, rank-8 idle remainder 576 - 1152 negative: a per-pair sink has no per-call declaration); strict mode rejects both declarations at SPEC MISMATCH expected.rank_mass: expected 1041408 got 1036800; a scalar re-window is rejected at SPEC MISMATCH providers.gauge44.complete_contract.

Named maintainer decisions

• MD-1, widened-window / common-sink contract change (the only decision that moves the moment ledger; consistent with Audit typed-echo to selected-centre recursive cost transfer #2500 that the typed contract alone does not imply fewer recursive children). Admitting the shared ledger (n_8 1,152 to 576, D 5,120 to 9,728) needs: the PR 386 sentence amended for one widened window with per-pair totals delivery; a per-pair sink declaration in the per-call receipt schema; the deficit literal replaced by the coverage identity; SPEC expected.rank_mass / expected.global_profile / expected_profile regeneration; and the moment engine's certify ceiling raised past about 0.0031852 or made adaptive. The chain's own moment oracle already accepts the shared endpoint.
• MD-2, supplier admission fields for typed windows. Declared fields for typed (frame-relabelled) window source rows, 8-row per-call windows and per-call totals delivery in the provider local receipt (today's 13 fields cannot express them), plus negative controls beside false_dirty_restoration in code/assembly_controls.py; a scalar re-window is rejected at the complete_contract / expected_local pins, so the admission route is a SPEC regeneration through the chain's discovery path with the new fields added.

Admission disposition

stage (pinned PR 1087 chain)	run against the typed-window schedule	result
strict T16 replay (11 stages)	fresh run on the retained local word	passes on the frozen data (control)
admission-controls (69)	fresh run on the frozen receipts	passes; no control can bind a typed-window declaration (no field exists)
supplier-control (12)	fresh run on the frozen receipts	passes; bit supplier untouched (derived deficit 10272)
verify_t16-style binding checks	final binding assertions on the fresh outputs	pass (roles, calls, prefix, row coefficient, margins)
invoice/assembly boundary, frozen receipts	assemble()	passes, kappa reproduced
invoice/assembly boundary, widened-window ledger	assemble() probes	rejected at named pins and literals (above)
moment certification, both engines	both ledgers at their own endpoints	accept at endpoint, reject at +1e-24 (four ledgers)
scalar re-window declaration	assemble() probe	rejected at complete_contract

Scope and limits

Admission-stage evidence only. The typed echo is not placed on the selected 144-call schedule: research/selected-centre-adjacent-pair-seed01fcb1 found no adjacent pair of that schedule sharing a physical owner among its eight centre source rows, so no placement exists without paid owner transport. The strict replay is the PR 1042 composition (kappa 0.001672188119205421), not the selected PR 1094 record; the selected record is unchanged. The probes' re-derivation of the invoice envelope mirrors code/assemble.py's formulas inside the probe harness only; the frozen files are untouched. Not checked: native emission, all-size/compiler, precision buckets, tape, analytic transfer, bit H1-H5. Not a multiplication theorem; admits no supplier.

Baseline. Canonical main moved during the session to 6bd33bc0b87a01e9c9d8b1c01576f3ddbcaee850 (priority revision 5 catalogue). The scientific baseline retained is the assignment-pinned 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2, intentional: every pinned chain input (chain code, engines, contract text, frozen replay archive, selected record) is byte-identical on the new main; only docs/research/agent-task-catalog.json changed, and this package pins the base revision byte-exactly for reproducibility.

Overlap (entry survey and refined passes 16:41-17:50 UTC, complete pagination; no novelty is claimed). Predecessors: #2464, #2595. Adjacent concurrent work distinguished, not overlapped: #2626 (fixed-reader fusion tariff), #2500 (typed contract vs recursive children), #2491 and #2543 (full dirty reader paired cycles), #2502 (carry error envelopes), #2158 (mutation coverage of the frozen invoice boundary: it mutates the frozen receipts to test guard coverage; this package declares the typed-window ledger against the same boundary and records where admission stops), Discussion 2576 (scoped full production replay of the unchanged chain).

Reproduce

python3 -I -B research/z4-typed-window-admission-kb1c/verify.py
python3 scripts/research_package.py --package research/z4-typed-window-admission-kb1c --check-only

Expected final line:

PASS_Z4_TYPED_WINDOW_ADMISSION_KB1C controls=42 admission=69 supplier=12 kappa=0.001672188119205421 D=5120/9728 no_new_kappa

42 controls cover the k=2 invoices, savings, witness rows and basis coverage, the pivot prices and the PR 2626 qualification, both coverage identities, the fresh replay verdict and its byte/semantic comparison against the pinned archive, all binding numbers, dirty-column counts, every boundary-contract location, the four moment accept/reject rows, the screen control, and all six assembly probes at their exact rejection points.

Credit

Mechanism and certified child family: PR 1391 / PR 1410 by Cinna Davis (1337isnot1337, with OpenAI Codex assistance), extending PR 1287 by Ross Richardson and PR 1227 by Sagar Thakrar (thakrarsagar); heterogeneous variant PR 1768 by Cinna Davis (1337isnot1337, with OpenAI Codex assistance) on one shared source; the two-bank variant PR 2098 by Aaron Wang (navinor). Priced instantiation and named obstruction: PR 2464; legal typed-window schedules and ledgers: PR 2595. Fixed-reader qualification: PR 2626 by petalas; typed-contract finding: PR 2500 (#2488). The pinned chain is Jeremy Kothe's PR 1087 package with Hao Yan's (FanPUYen) T16 controller and the credited bit/complex lineage, orchestrated by Douglas Colkitt; frozen receipts: research/community-frontier1042-audit/selected-replay.tar.gz. Immutable-window sentence: the bit-geometry artifact PR386-BIT-GEOMETRY.txt in PR 386 by Zhihao Chen (jacklightChen, with OpenAI Codex assistance), whose local E8 bit word is Dylan Crow's from PR 370 (with Anthropic Claude assistance), pinned in research/split-centres-global; adjacent-pair owner census: research/selected-centre-adjacent-pair-seed01fcb1; published D/E: research/selected-suppliers-de-diagnostic (Navinor); pricing model: research/complex-local-word-ceiling (Douglas Colkitt).

New package material is Apache-2.0. Prepared by Aaron Wang with help from opencode (substantial AI assistance in derivation, implementation, checking and prose). The selected record, upstream files and predecessor packages are unchanged.