Research: audit already removed QQ16 centres and actual outer consumer boundaries

[Clear1nsight]

The proposed centre-retirement task has a changed source premise: all eight selected H8 centre children are already absent from the actual QQ16 core. joint-query/source_data.py removes their phase and inserts a paid dirty-state correction. This audit independently traces their original lifetimes, regenerates both corrected words byte-for-byte and pins two real outer consumers of PR3952's reduced F17 interface. It establishes precise obstructions, no new supplier or kappa.

The inherited correction contains 580 ADDs and 16 CLEARs. After subtracting the removed scalar phase, net scalar increases are 3,372 forward / 3,324 backward; each corrected core has scalar allowance 202,138 and native rank mass 5,522. Actual F0 and B1_2 have no H8 and add respectively eight and 24 H1 lines. Original C buffers clear by event12758 before call exit; their effects have been applied to L. Existing removal credit belongs to the QQ16/source lineage, not this audit.

Two pinned boundaries explain why the 96 earlier y1_U outputs do not close the caller:

• Actual ordinary bridge before F2: the hosted physical prefix includes a pair1 idle sweep between B1 and the bridge. Eight independent, reachable integer inputs, constant in both owner and payload, give the same 96 projected zeros but different bridge outputs. The bridge therefore cannot factor through that projection alone. The template still names its original gauge44 provider; this audits the real continuation proposed for a QQ16 substitution, not an admitted complete substitution.
• PR2707 full F0-inverse suffix: checkpoint2740 requires all224 F17 x1/y1 values. The projection lacks112 x1 and16 selected y1 fields; event2741 already requires absent Z0. Exact readsets, reachable collisions and all128 missing boundary directions are checked. Missing fields are not a lower bound of128 new families: current borrowed X/Y remain available to a redesigned paid consumer.

The retained complete recovery fallback is fully priced: 336 H1,672 wrappers,464 maps,1,808 unit ADDs,896 quarter ADDs,312 CLEARs; scalar allowance25,840; private16. With H8 absent it still fails the unchanged conditional pair screen: [1.000097677880011248,1.000097677880011249] >1. This prices an existing feasible fallback, not an optimality theorem. The reduced112-H1 endpoint passes that relaxation, so the obstruction is composition and access, not its local algebra.

Independent rational arithmetic reproduces the old fixed-total two/three-H8 threshold, while distinguishing reallocation into ideal idle work from literal deletion that changes total mass/centre. Neither grants another retirement credit. Family stock, invocation counts and coverage must be proved for a changed consumer before transporting the screen. Physical movement/placement, cleanup and return, precision/dirty-helper ranges and both all-depth supplier caps remain open.

Only research/qq16-centre-lifetime-boundary-audit/ is added. PROOF.md contains the argument and limits; RESULT.json retains the full checked receipt. SOURCES.json pins twelve canonical inputs. PR3952's WORD is retained as authenticated data, not an executable dependency; replay imports no predecessor executable. PREDECESSORS.json records exact heads, blobs and reuse. Frozen science and selected frontier are unchanged.

Validation on Linux CPython3.12.3, standard library:

python3 -I -Werror -B research/qq16-centre-lifetime-boundary-audit/verify.py
python3 -I -O -Werror -B research/qq16-centre-lifetime-boundary-audit/test_verify.py
python3 -B scripts/research_package.py --package research/qq16-centre-lifetime-boundary-audit

Normal/optimized receipts are byte-identical. Registered replay passes; seven integrity/export groups pass, including read-only Git-free replay, pre-execution source authentication, retained-word corruption and typed whole-receipt rejection. Exact evidence covers all eight C lifetimes/frames, corrected-word equality,712 scalar columns, helper/private cleanup, actual bridge collisions, full recovery input census/invoice and outward rational intervals. Published-byte/source-history preflight passes for 28 package/input files at head82c812c45fbc2bc2751e5a0d02a3631af25c3cba; whitespace check passes. VERIFICATION.json lists unrun selected cold replay, huge native kernels, physical tape execution, all-depth supplier proofs and human/formal review. Hosted run39214876589 passed planning and package replay at exact head82c812c45fbc2bc2751e5a0d02a3631af25c3cba. Logs bind its merge into base20ecf289, CPython3.12.15, seven tests in31.224 seconds, registered completion and unchanged committed inputs.

Explicit assigned continuation of #3952 head996ff698d3efd0bdce606e240f3b00e6ab91e7fc, following #3946/#3908. Canonical main20ecf2891d80b5acf18d980b2164fb44ea7faa90 and priority revision6 were refreshed at start, source-impact checkpoint and publication, unchanged. Scope refined when exact sources showed no surviving centre target. Three user-requested agents independently audited provenance, recovery boundary and budget, then cross-reviewed one another; a separate agent also reviewed the root bridge witnesses and integrated scope. REVIEWS.json records limits; no independent human review. Targeted live PR search plus the repository helper covered adjacent work partially: no gh/complete remote snapshots, hence no novelty guarantee. Canonical claim.

Next bounded step: choose one exact paid composition boundary. Either reconstruct/encode the128 missing full-recovery checkpoint values from current X/Y more cheaply than the336-H1 fallback, or let the real bridge consume a changed encoding while identifying source reads/operations that genuinely disappear. Derive arbitrary-array equations, frames, lifetime/return and a full bill; establish integer invocation/family coverage before using the budget. Stop on a missing access/cleanup contract or a failing complete invoice. Further centre deletion or isolated precision work is not justified by this result. TASK.json retains original card, checkpoints, first steps and blockers.

Credit Ali Romman's QQ16/source lineage, centre-encoding/cover-algebra/residual-syndrome/native/interface lanes and retained gauge44/T16 notices for the existing removal; ywz65 #2512/#2707, Arham Tahir #2681, alephnan #2346, thakrarsagar #2470,1337isnot1337 #3908,dcharb78 #3181,#3946/#3952,Douglas Colkitt,Sagar Thakrar and coverage-ceiling contributors. New work by Chris R / Clear1nsight with substantial OpenAI assistance and three requested AI agents. Apache-2.0; predecessor notices unchanged.

[Clear1nsight]

FINDING / DONE — three distinct agent-owned research items published as separate PRs. Bounded research and hosted verification complete.

Item	Artifact / exact head	Established result
Actual invocation/coverage audit	#3976,137c557bbd5fb9f2221457fa400e3ed439481922	16 proposed first-B1 insertion sites/144 local calls=1/9; nine-word meanH1=2412+136/9. Conditional336 module passes;553 passes/554fails. Old pair-rate rejection cannot transport.
Paid-query bridge composition	#3977,c8e72091d379972d510a409ede15eaf44eb7b6be	104-field h+a boundary closes the information gap; serial208H1,416wrappers,400maps,15040scalar. Exact fusion leaves existing96Y-bridge ADDs, with no additional source work removed.
Full recovery fusion	#3985,400df7fca8f60c19bf163c77564fe3fbe9154b04	Same224 original F17 outputs:336→224H1,672→448wrappers,464→352maps; scalar25840→27184 and112 additional exclusive full-cover temporaries, all zero/identity returned.

The incidence result concerns the frozen1042 hosted template; all144 source calls remain gauge44, with zero literal3952 consumer invocations. One-stage1-module is an insertion proposal, not sixteen proved demands or current1094 supplier admission. W896/9,centre64 and ideal idle completion are explicit conditional assumptions; no productive-family expansion or displaced-idle credit is proved. PR3339 already supplied the136 count and idle-additivity warning; the new ledger/source-rate consequence is distinct and credited.

The bridge result is a local changed common-label contract. Direct Y subtraction costs768scalar, or1008with the same120workspace clears, under that contract. It does not establish completeQQ16 inter-pair placement or a universal consumer lower bound. Full recovery preserves all224 borrowed current inputs, uses no early records and keeps private peak16. A nominal120-H1 borrowed-Y bypass failed the full endpoint atZ0 because it omitted earlier inverse-prefix corrections; it is explicitly not a construction.

Validation: each package has exact normal/optimized replay, semantic/integrity controls, read-only Git-free export, registered metadata/replay and published-byte/source-history preflight. Separate sibling AI cross-review and coordinating review found no substantive blocker in declared scope; exact reviewed hashes and later hardening are recorded. No independent human/formal review, full selected cold replay or physical supplier admission is claimed.

Hosted3976 run39218396413 PASS; hosted3977 run39218677372 PASS. Exact head/base, CPython3.12.15, seven tests each, registered completion and unchanged committed inputs verified. Hosted3985 run39219936311 PASS at head400df7fc/base20ecf289 on CPython3.12.15: primary verifier, five boundary groups, normal/optimized/read-only export and final committed-input diff all passed. All three worktrees are clean at their GitHub-created heads.

Next useful decision: bind one actual provider/consumer replacement at an enumerated first-B1 site, then identify actual displaced work and productive-family coverage. For the224-H1 route, map the112 extra temporary lifetimes to paid owned/dead or restored dirty storage and price movement, numerical/dirty-range changes and both supplier caps. For thebridge, require a changed producer deleting named work or an actual second consumer sharing h/a; do not continue serial decode/reconstitution alone. Stop on missing read/return/coverage or a failing complete bill. A favorable relaxation is not permission for another unbound decoder/precision campaign.

Main20ecf289 and priority6 unchanged at entry,07:40,07:45,07:53 and08:03 publication checkpoints. Targeted live PR/Discussion overlap refreshed; helper remote coverage remains partial without gh. Frozen sources, selected frontier and others' PRs unchanged. Credit Ali Romman/QQ16, ywz65/2512/2707, Arham Tahir/2681, Nick Petalas/3339,3952/3963 and retained geometry/native/gauge44/T16/owner/coverage lineage. Chris R / Clear1nsight with substantial OpenAI assistance and three separately assigned AI agents.

Original advance claim and bounded expansion decision

CLAIM — three distinct paid-consumer follow-ups, user-assigned parallel session (2026-10-19 07:25:54–08:25 UTC; final quarter reserved for verification/handoff).

Canonical main20ecf2891d80b5acf18d980b2164fb44ea7faa90, priority6, entry guidance and family gate freshly read; baseline bindings pass. Predecessor3963head82c812c45fbc2bc2751e5a0d02a3631af25c3cba remains open/unchanged and independently source-bound locally with hosted replay, not admitted as a supplier. No new centre retirement: QQ16's H8 are already absent.

Three agents own different objects and separate branches/packages/PRs if useful:

1. Paid bridge composition: actual physical source_event3 after F0/B1 and pair1 idle. Change the consumer/query boundary, derive arbitrary-array identities and frames, identify actual source operations eliminated, and price access/restoration. Stop on a failing complete bill or exact missing module.
2. Full-recovery fusion: use retained current X/Y to supply the PR2707 checkpoint or fuse the suffix more cheaply than336H1. Test an explicit changed encoding/circuit, not another missing128-field census. Stop on a source-bound obstruction or unfavorable complete allowance.
3. Actual invocation/coverage: reconstruct integer consumer incidence and productive-family lifetimes on the real hosted schedule. Test whether896/9 and centre64 legally transfer; do not assume a hypothetical4.5 consumers/ninecall schedule.

Expansion decision: narrow, bounded contract tests/audits, not a supplier implementation campaign. Known/proved starting evidence is3963's source correction, two interface collisions and retained fallback invoice; reduced112-H1 endpoint is constructively priced locally. New source access/frame/lifetime/return, global placement/movement, precision/dirty-range, coverage and integer invocation allowances remain unknown until an item proves them. Unknown costs are not zero. A favorable screen alone will not expand construction. Each item must name real replaced work or establish a new precise obstruction; one bounded pivot maximum.

Overlap: fresh3963head/comments, targeted live QQ16 bridge/recovery/invocation PR searches and active QQ16 Discussion index checked. Nearest relevant work includes3952,3946,2707,2681,2512,2470,3394,3652 and output-bank/consumer threads2617/2657; exact reused inputs must be independently checked/pinned. Repository helper cannot refresh complete remote snapshots without gh: coverage partial, not novelty proof. Original handoff1166 and start-here1163 conventions retained. Each agent refreshes main/overlap before major pivot and submission, preserves frozen inputs, records original task card and source attribution.

No selected frontier changes, merges or closures. Separate AI agents and root coordination constitute substantial OpenAI assistance, not independent human/formal review. Current status and eventual artifact links will update this single canonical comment; handoff hub gets a concise pointer.