[Clear1nsight]

Change and mathematical scope

Independent audit of Dillon's #5402 provider at exact head 3348fccea855082786d5f0fc5558b73d4974e745, its literal ordinary work and a conditional closure-budget screen. A companion bounded review covers Ali Romman's #5376 all-width Clifford criterion and certificate boundary at 848cdce9003262c4b6e02ec4558f302e4fae56ad. No complete caller or new κ is admitted.

The independent provider interpreter passes all 632 arbitrary original inputs plus nine arbitrary owned entries, derives the unrestricted source identities and 122-owner basis rather than importing the candidate evaluator, verifies all 1,240 physical phase equations, and returns all nine owned fields to zero. Ten controls reject actual instruction, basis, frame, absolute-phase, operator-order, false-credit and cache-content defects. The additional cache SCALE−1 control stays in the same physical chart and fails specifically because a later zero-retag input becomes nonzero.

The accepted endpoint is X=Dx−Ry+RWh, Y=Dy, H=D(h−Zy), with ZW=14I8 and WZ=I−C, preserving noncommuting order and exterior scalar coefficients at every positive width. The eight H outputs remain coupled and earn zero autonomous Gaussian credit. Four hundred other helpers retain their actual autonomous contract.

The paid profile already equals the validated #5398/#5400 comparison. The deletion of eight H8 and eight H1 calls is relative to the original raw source and is not another paid saving over that comparison. For a SPECIFIED hypothetical closure into the same eighteen-call p16 comparator, grant eight extra autonomous H8 targets per provider solely as a favorable assumption. At κ=.00302728823, all144 hypothetical targets have moment credit about1144.77085846746. Strict incremental aggregate homogeneous cleanup budgets are1144H1,143H8 or128H9, with all other jobs unchanged. 144 paid H8 exactly cancels144 target H8 rankwise. This is a necessary conditional screen, not a construction, funded endpoint or universal coupled-controller ceiling. Every extra paid/lost-target/main/idle term and unknown ordinary cost remains explicit.

Literal counts are17,917ADD,1,204SCALE,16COPY,40CLEAR,eightSWAP and464monomial maps per provider. The corrected exact scalar tariff is175,022 units; the earlier independent checker inverted an already-final boundary coefficient twice and gave a conservative175,038. Its original artifacts are retained as version-scoped diagnostics. Counts, recursive profiles and moment budgets are unchanged; v3 uses the emitted factor directly. Monomial realization and consumer costs remain separate.

A fresh all-width local error proof derives exponent18,996 in the maximum-over-fields Euclidean norm, independently matching the author's stated geometry, and19,012 in the joint-field Euclidean norm. For either geometry every contiguous subinterval/suffix has norm at most2^L, giving ||e_out|| <= 2^L (||e_in|| + sum ||local_residual_j||) at one fixed width. No coordinate-to-field conversion, recursive precision, tape/storage or compiler funding is hidden in that local statement.

The companion #5376 written-proof review covers trace separation, eight-copy invariant descent, zero/one-axis exceptions, sharp one-axis cutoff and the stated stability domain. Its independent oracle checks24programs/64directtensorwidths,17literalwitnesses,68forgedwitness rejections and a separately constructed96-matrix group/rank certificate. Submitted modules are comparison targets only. The negative witness checker does not independently accept a positive compact receipt, and shared word expansion remains a documented trust boundary. This package does not admit the complete #5376 provider,7,000 classes or30,328 fusion census.

Task and overlap

Donor-assigned gpt-6.1-sol high swarm. Task independent-moment, priority revision7, changed to independent partial-provider/conditional-budget review with a bounded verification-foundation companion. Canonical main 1ec1119972c9aefe42525172d07b332fcff81cb4 was freshly fetched before research and submission. Only research/coupled-provider-independent-audit/ is added.

#5405's current-owned-profile controller ceiling is not repeated or transferred to this coupled contract. #5400's κ=.00302728823 is a frozen comparison, not rerun admission. The live Dillon handoff releases the caller scope; its private continuation is unvalidated here. The proposed A S A = Q S control still needs concrete ordered operators, charts and helper-tail implementation. No competing caller or new source construction is claimed.

#5376 retains its older scientific branch base; newer canonical provenance pins are authenticated without claiming canonical main is its ancestor. Original raw results remain unchanged. A separately derived algebra replay expectation adds only the authenticated adjacent same-chart-control script hash, because that file was added after the original root receipt; source_guard reconstructs and constrains that one-row metadata change. Mathematical comparisons remain exact.

Validation

Python3.12, standard library only:

python3 -B scripts/research_package.py --package research/coupled-provider-independent-audit
python3 -B scripts/check_research_commit.py --package research/coupled-provider-independent-audit --check-source-history
python3 -B scripts/export_research_package.py export --package research/coupled-provider-independent-audit --ref HEAD --output /tmp/coupled-provider-audit.tar
python3 -B scripts/export_research_package.py check /tmp/coupled-provider-audit.tar

Tested publication head b62252c61113d2863b36ce2419aa81a07f393c42, tree 7917f77040070a23333e715441e9445583d7d336:318committed files,316bindings plus manifest/verifier; working/committed/export bytes match. Canonical-only source-history and offline side-head origin proofs pass.

Fresh serial registered normal/optimized wrappers PASS in75.694s,238,108,672bytes sampled peak descendant RSS. Both modes deliberately compile frozen scientific source at optimize=0, preserving assertion-based checks; this is not an unprotected direct-O claim. Adapter falseassert tests for main/local imports and rogue-module/bytecode/falseH8/κ controls pass. Every complete stable mathematical result is compared; only named runtime fields are excluded.

Five fresh stages are algebra, same-chart control, corrected budget, local envelope and bounded Clifford oracle. The separate unmodified author #5402 registered pair PASS51.350s, with all16runtime inputs unchanged; it is authenticated frozen evidence and is not rerun by package CI. Full #5376 registered/provider/fusion replays were not run here.

Official export and read-only Git-free normal/optimized pair PASS75.390s,221,880,320bytes sampled RSS. Inputs remain0444/0555, PATH excludes Git, all318 source bytes unchanged. Export SHA256 7ac9d8c8276ad68de205ee0fa699e6c9aaa72149d1846a46f88a28d143200173. Hosted Registered research packages run 40041031723 PASS: onboarding/selection, the explicit bounded package replay, and committed-input immutability all succeed at this exact head. Hosted CI remains separate from scientific admission. No Lean, Docker, proof-kernel or independent human acceptance is claimed; inherited physical/compiler/precision/tape/analytic foundations and selected result remain unchanged.

Credit

Dillon/dillona supplies #5402 and #5388; Ali Romman supplies #5383's caches, #5398's caller and #5376's Clifford proof/implementation. Cao Yifan supplies #4610, Dugongue the center-free code, Cinna Davis the retained helper/controller lineage, and zxv, Jacob Alan Sussman, ToikY and all native/Clifford/compiler predecessors retain credit. The Clifford source notice preserves Rohan Arun, Tekalign, Jeremy Kothe and retained contributors. Chris R/Clear1nsight supplies this independent audit with substantial OpenAI Codex assistance through gpt-6.1-sol high agents. Exact notices retained; Apache-2.0.

Open next questions

Supply one exact caller or typed recursive closure that handles all224 main and eight coupled helper outputs while retaining400autonomous helpers, observation cleanup and all original inputs. Bind the actual ordered control/helper-tail words, then apply the conditional budget with every added Gaussian job and lost target. A passing necessary screen still needs complete precision, stock, movement, compiler and outer funding. Separately, independent compact positive certificates or broader native oracle coverage can strengthen the Clifford verification boundary.