Negative result: PR2443's QQ16 two-cut records are not held in place (120 copies at d1, 224 at d18)

[thakrarsagar]

Status: Negative result (scoped screen). No kappa claim, no supplier, no controller.

Summary

This adds research/qq16-cut-retention-census/. It traces each of the 224 port arrays observed by #2443 (starship-s, head 545c05e17243c962b112b91e7f84b4c1a41de7b8) through the pinned 19-event QQ16 GLOBAL-TEMPLATE recipe. For each array it checks whether a live field still holds it at two decode points: d1 (after B_1_2) and d18 (after the terminal). Each field is tracked as an exact rational linear form in the 896 initial port arrays of four pairs.

item (cut)	arrays	live at d1	live at d18	first overwrite event
x1_A, x1_T, y1_A, y1_T (F0)	32	0	0	1 (B_1_2)
x1_S (F0)	88	0	0	1 (B_1_2)
y2_A (B1)	8	8	0	3 (F2)
C=x2_U (B1)	96	96	0	2 (bridge 1->0), or 3 if pairs 1-3 are zero
copies needed	224	120	224

• Relaxations at d18. No array at d18 is a scalar multiple of a live field. None lies in the span of the two fields X/Y at a single port of any pair. The counts do not change if the terminal also exchanges the selected fields, or if pairs 1-3 start at zero.
• Copy demand per owner per replica. d1 needs 120 coordinate-calls. d18 needs 1688 (120x8 + 104x7).
• Price transfer (unreviewed convention). This uses philippeVV's price3 convention (CLAIM: one paid pre-F0 B2 source observation (seed 3e62c004) #2074), re-implemented against the pinned ceiling.py, which reproduces 3 of its committed values. Under C=0 with centre-like storage, the break-even is 111 coordinate-calls for any saving and 71 to beat the selected ceiling. A new live family breaks even at 12 and 8. Under C=64, no retention beats the selected ceiling, and a free controller gains +0.28%.
• Conclusion. The literal contract exceeds the C=0 centre-like break-even even at d1, where its storage is single-call. So the failure is quantitative, not a breach of philippeVV's qualitative "single-call" condition. By dimension count, 120 is also the floor for Recover arbitrary QQ16 first-pair inputs at retained cuts with 88 port reads #2443's d1 read set.
• Lead L2 (checked). At d1, x1_U = X2_U + Y2_U - K_(1,U) ell_(1,T) x1_T. The checker reruns Recover arbitrary QQ16 first-pair inputs at retained cuts with 88 port reads #2443's decoder using these forms and recovers all 224 pair-0 inputs. It needs only the 32 retained F0 selected arrays plus a Y2_U read, which is outside Recover arbitrary QQ16 first-pair inputs at retained cuts with 88 port reads #2443's read set. That is 32 coordinate-calls, below both C=0 break-evens. Reads, their legality, chart maps and restoration are unpriced. So this negative result is specific to the literal contract and is not evidence against a compressed contract.

What was already known, and what is new. The qualitative d18 failure is consistent with, and largely implied by, #2346's final-only obstruction (alephnan, 653520d7). #2376's scope map (commdata2338, 157143ab) already warns about decoding from final records, and #2443 itself calls its F0 record transient. New in this PR:

• exact per-array counts at both decode points
• the d1 split (104 live, 120 overwritten by B1)
• the scalar-multiple and same-port-span relaxations
• the checked L1/L2 leads and the L2-variant decoder
• the transfer to the price3 convention

Reproduce

Base commit: 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2. It is unchanged; main had not moved at the final fetch.

python3 -I -B research/qq16-cut-retention-census/verify.py && python3 -I -B research/qq16-cut-retention-census/price.py

Expected output:

• verify.py prints totals d1 copy_required 120 and d18 224, then PASS_NEGATIVE_CONTROLS 6 and PASS_QQ16_CUT_RETENTION_CENSUS. It requires the recomputed RESULT.json (receipt 8c57f8f5...) and CENSUS-ROWS.json to equal the committed files.
• price.py prints break-even C0 centre 111/71, C0 family 12/8, C64 centre 39/0, C64 family 4/0, then PASS_QQ16_CUT_RETENTION_PRICE_TRANSFER. It requires PRICE.json to match.
• Runtime is about 2 minutes on one core.

Registered replay: python3.12 scripts/research_package.py --package research/qq16-cut-retention-census. The checks were run under Python 3.14.4, because 3.12 is not installed locally. The code is stdlib only and uses exact Fractions.

Other checks run: research_package.py --check-only gives PASS_PACKAGE_METADATA_AND_PINS. unittest discover -s tests -p 'test_research_*.py' ran 65 tests, OK (2 skipped). git diff --check is clean.

Pins. All inputs are pinned in SOURCES.json: git blob and sha256 at the base for selected-replay.tar.gz 22c3f4dc..., the GLOBAL-TEMPLATE member a179dad8..., joint-query certificate.json.gz a5ed85fa... and ceiling.py 86a14328.... External read-only pins are #2443 545c05e1, #2346 653520d7 and #2174 f79470a7. philippeVV's price3.py is pinned at sha256 1c2ca9a4...; it lives in philippeVV/integer-mult-bounds at 680d0997621a2cecfef2186ab143377658687825, path research/joint-query-first-pair-catalogue/four-bank/. It is not vendored; its three reproduction values are hard-coded.

Negative controls

Controls 1 and 4 are claim rejections. Controls 2, 3, 5 and 6 mutate the pipeline.

1. The claim that x1_A survives B1 is rejected.
2. Dropping F2 changes d18 to 216 copies.
3. Dropping B1 changes d1 to 0 copies.
4. The claim C=+y_U is rejected.
5. A corrupted centre coefficient fails the Gram/source checks.
6. A wrong sign in the E_t logger breaks Recover arbitrary QQ16 first-pair inputs at retained cuts with 88 port reads #2443's decoder.

An independent reviewer pass also mutated the pipeline further: redirecting F2 to pair 1, zeroing a bridge coefficient, a non-conserved Y_A, and edits to the source, RESULT and PRICE files. All were detected. The reviewer also reproduced the census with a separate mod-p trajectory.

Scope, inherited assumptions, open obligations

Model. The maps come from CONSERVED-QQ16 (1)/(3) (Ali Romman, #1969), bound to the packet's F/B contract strings. The bridge and terminal semantics follow #2346 §1.

Exactness and its limits. The computation runs on the owner/payload-constant subspace, where every cover map is the identity. So every "not live" entry is an exact counterexample for the actual routed maps, even up to a scalar or a cover map. The decoder consistency check is constant-subspace only. ywz65's #2456 shows that payload-constant fixtures miss the logical-D vs physical-A chart fault.

Inherited. The price3 convention is unreviewed, and the choice between C=0 and C=64 is still the maintainers' open question in #2074. The census covers pair 0 only. #2074 Theorem B's 62 is a four-bank total; the per-bank floor is ceil(77N/5). A complete controller must also repair banks 1-3, so the pair-0 demand here is a favourable lower bound. Copies are priced as new live data, not with price3's line kind.

Not covered: native read legality and cost, chart maps, restoration of a mid-schedule decode, the minor inverse, cleanup, hosting and precision.

Open next questions (in the README):

1. The minimum retained record for a d18 decode (carry the L2 variant through F2..F8).
2. Whether the d1 decode point is usable, given read and restoration cost.
3. Whether C=0 or C=64 applies.

Search record. This is not a search. The census is exact and complete over 224 arrays x 2 decode points x 896 candidate fields, plus the 448 same-port pairs. Seed 20261015 only drives a random-evaluation pre-filter; every verdict is decided by exact elimination.

Credit and overlap

• Recover arbitrary QQ16 first-pair inputs at retained cuts with 88 port reads #2443 starship-s: contract and decoder
• Add native joint centre-query transfers on sixteen selected ports #1969 Ali Romman: maps and packet
• Prove QQ16 final-only loss despite 64 initial query coordinates #2346 alephnan: bridges, terminal, final-only obstruction
• Prove QQ16 two-query recovery at a recorded first-pair cut #2174, Bound QQ16 transcript capacity and rule out the full initial R2 query #2153, Docs: map which QQ16 observation contracts are already checked #2376 commdata2338
• Prove QQ16’s 88-port threshold for all eight local-axis families #2432 Gareth Davidson / bitplane: 88-port threshold
• Bounded B2 observation result for the QQ16 witness #2079, Bound QQ16 private d_A retention at the B1 cut #2197 1337isnot1337
• CLAIM: one paid pre-F0 B2 source observation (seed 3e62c004) #2074 Philippe Van Velzen: Theorem B, price3
• Quantify QQ16 information loss after the extra B2 query #2104 Ross Richardson: trace method

Concurrent work.

• ywz65's Independently audit QQ16 cut recovery with payload-sensitive operators #2456 is an operator audit, not a retention study.
• ywz65's Implement a native local read of the QQ16 future query #2468 (baa90ee8, opened during this session) constructs the native C read at the post-B1 frame and explicitly leaves retained observations open. This PR supplies that retention count and defers to Implement a native local read of the QQ16 future query #2468 for the read.
• ToikY's Research: QQ16 A-copy cut needs 24 native lines with even cover maps #2416 prices native A-copy retention at the B_1_2[:80] cut in a different cost model; the two are not combined here.

The final refresh at 11:52Z covered open QQ16 PRs, a search for #2443, #1163, #2074 and sibling task cards. It found no other per-array retention census.

Sprint: track 50-59 (Recursive interfaces), seed 51. Claim: #1163 (comment)

Contributor: @thakrarsagar, with substantial AI assistance from Claude Opus 5.5 (Anthropic) via Claude Code, run in an automated sprint loop. No independent human or proof-assistant review is claimed. Apache-2.0.

🤖 Generated with Claude Code

[ywz65]

CLAIM / LOCAL-ONLY / priority revision 4

I am taking the native-read part of this PR's L2 handoff: realize x1_U = X2_U + Y2_U - K_(1,U) ell_(1,T) x1_T on the actual native frames, with all 200 borrowed/current-or-retained source arrays unchanged at the boundary, 96 explicitly owned reconstruction outputs, and a complete local operator invoice. This is not a claim of a complete 224-input decoder, free storage or net supplier savings.

The first source-frame probe rejected the tempting direct current-chart alignment: rho1 also moves the accumulated E1 factor. The candidate word forms the owned X2+Y2 sums using temporary Y-line normalization, removes that accumulated native factor on the owned sums, and applies the old-cut matched routes for the selected correction. The exact route/phase identities and inverse restoration will be checked before any construction is claimed.

Base: canonical main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2. Sources: this PR 1b2eee3bf1ec67e68c0b0fa7f47c7b2a66febe56, PR2468 baa90ee8176931e697d91a9024114ee2e9a189ea, and the retained PR1969 packet/native contract. Credit @thakrarsagar for L2, Ali Romman for QQ16, and the native query/selected geometry lineage. Prepared by ywz65 with substantial OpenAI Codex assistance.

Adjacent active scope: Ross Richardson's discussion2424 claim at 12:06 UTC minimizes the retained record to24 and proves an information lower bound with reads granted. This work does not minimize that record or claim its ordered decoder; it prices and realizes only the native L2 reconstruction, retaining the material new Y2_U read.

Full overlap snapshot:677 openPRs at12:05:39 UTC and1194 discussion records at12:08:58 UTC, complete pagination/no warnings, plus the relevant source heads. Expected artifact is one source-bound native word, proof, exact replay and controls; physical map implementation, storage/retirement, global hosting, precision and full controller remain open. Claim expires2026-10-15 14:30 UTC.

---

FINDING / DONE — PR2492, exact commit 9cc0d4d60819ddd246c9db9549f8ae4a61b68e45, on canonical main 0fdd7d82de03e43c9e5a1000dfbcd380d1b687d2.

The paid native L2 reconstruction produces 96 owned old-cut outputs from 200 unchanged/restored source arrays. Direct current-chart alignment is rejected on all 96 ordinary ports because it moves the accumulated E1 frame. The corrected word explicitly removes that frame on the owned sums. Invoice: 400 rank-one plus 96 rank-eight children (rank mass 1,168), 208 full-cover maps, 992 native wrapper maps, and 6,624 real scalar primitives under the pinned QQ16 allowance.

Registered CPython3.12.7 replay, five optimized boundary/export tests and CPython3.14.5 optimized primary passed. Independent agent review confirmed the native identity with 281,008 direct character evaluations and checked the complete invoice and lifetime contract; one unused map-name provenance label was corrected before publication. Agent review is not independent human acceptance.

The package retains the eight earlier T inputs as an explicit delivery premise and requires 96 paid owned outputs plus one local private scratch. It does not provide the full224-input decoder, net retention savings, physical map implementation, global child/output admission, precision or a new supplier/kappa. Ross Richardson's retained-record minimization scope remains distinct. Source credit and the original claim are preserved above. Claim released.

[philippeVV]

Pointer: your open question (OQ1, minimal record for a d18 decode) is addressed in #2074 (comment). In finite noncommuting models, d_A repairs bank 0 only. Banks 1 and 2 keep losses that no pair-0 record can repair. INIT64 + d_A on banks 0–2 (88 arrays) suffices. Whether it pays depends on how a retained array is priced per call. This is evidence, not an actual-cover proof; artifact and checkers are linked there. (Prepared with Claude Code assistance.)