DLamdata · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026 · Nov 3, 2026
diff --git a/research/pr6551-retirement-audit/ACCOUNTING.md b/research/pr6551-retirement-audit/ACCOUNTING.md
@@ -0,0 +1,83 @@
+# Independent accounting of the retirement tree rebalance
+
+At candidate PR6551 head `11ca568cd1a16776ade02b4df09ea04a784cc674`, relative to predecessor PR6545 head `775722aa94a5e85b05d1d704d7dca10cbf016f0c`, direct counts of the literal callers give
+
+```
+ΔT = −104 H16 + 104 H40 + 208 H41 − 208 H57,
+ΔP = −104 H16 + 312 H40 − 208 H56,
+Δ(T−P) = 208[(H41−H40) − (H57−H56)].
+```
+
+The caller counts agree with the predecessor's claimed change. The checker does not import or execute either author's invoice, pricing, constructor, native or caller code, and does not read their RESULT.json files. The first direct event/endpoint count was derived before reading the candidate proof's complete signed invoice. The written physical semantics and literal data are shared inputs; this is independent accounting implementation, not independent derivation of the native tariff.
+
+## Two representations and the exact count
+
+For one port, a declared endpoint `q^a ∏ D_s` has positive Gaussian rank `|a|+8·#{s}` in this caller. There are no stage-zero targets. The central exponent −1 is normalized by an ordinary q² translation; zero means no central axis. Each binary whole IDLE pays for two jobs of rank eight times its stage-set cardinality, IDLE_ONE pays one, and QUERY pays one rank-one job. Each LOCAL expands into the shared native subprogram.
+
+The checker first traverses every literal child tree, assigning the eighteen fresh stages to leaves. It obtains internal-node sizes `{2:8,3:2,4:2,7:2}` before and `{2:8,3:2,4:2,5:1,7:1}` after. The entire root outside `children` and `child_sizes` is identical. The first two children retain sizes seven and two; C,D change from two,seven to five,four. The masks AB and CD still contain nine stages.
+
+The root-mask count, the two productive outputs of every internal node, and the independent retirement split give an expected histogram. It must equal a separate count of all 35 serialized endpoint rows. The checker also verifies the actual retired field 4 and fresh suffix field 34, so exchanging their targets cannot hide behind an unchanged histogram.
+
+Separately, all literal events are classified. The fourteen binary IDLE stage sets must equal the actual tree-node stage sets, including multiplicity. The two IDLE_ONE stage sets must equal children B and C. All seventeen binary/root thin queries and all eighteen ordered LOCAL stage calls must be present. Unrecognized event kinds fail. Q_POWER exponents must be ±2, the inherited ordinary translation case.
+
+Per port, excluding the identical native contribution:
+
+| Rank | Old productive | New productive | Old paid | New paid |
+|---|---:|---:|---:|---:|
+| 1 | 2 | 2 | 17 | 17 |
+| 16 | 1 | 0 | 18 | 17 |
+| 17 | 17 | 17 | 0 | 0 |
+| 24 | 0 | 0 | 4 | 4 |
+| 25 | 4 | 4 | 0 | 0 |
+| 32 | 0 | 0 | 4 | 4 |
+| 33 | 4 | 4 | 0 | 0 |
+| 40 | 0 | 1 | 0 | 3 |
+| 41 | 0 | 2 | 0 | 0 |
+| 56 | 0 | 0 | 4 | 2 |
+| 57 | 4 | 2 | 0 | 0 |
+| 73 | 2 | 2 | 0 | 0 |
+| 145 | 1 | 1 | 0 | 0 |
+
+The rank-40 paid coefficient three consists of two binary-node idles and the third child's additional root idle. Dropping that last job would produce a false improvement.
+
+The exact retained native archive contains 104 distinct ports, 770 public fields (562 productive helpers plus 208 main fields), and 3,456 MOVE events. The candidate source archive contains a native archive whose SHA256 matches the independently obtained exact PR6545 native archive, plus byte-identical copies of the predecessor caller and retirement witness; its external WITNESS.json is also identical. There are eighteen native invocations in both callers. Thus the common native paid and helper terms cancel. On the retained structural contract, both complete invoices have 13,756 productive fields and 67,096 paid jobs. Both target and paid rank mass change by −832; their difference is unchanged. Two extra Q_POWER translations are explicitly reported as ordinary events, rather than silently dropped from the event inventory.
+
+## Strict gain for every 0<t<1
+
+Set H_r=r^t. For 0<t<1, `f'(x)=t x^(t−1)` is strictly decreasing on positive x. For every s in [0,1], `f'(40+s)>f'(56+s)`. Integrating gives `41^t−40^t > 57^t−56^t`. Multiplication by 208 proves strict positive Δ(T−P). The difference is zero at t=0 and t=1; neither endpoint is claimed strictly positive. This proof establishes a complete recursive Gaussian invoice improvement in the retained model, not the new numerical κ or physical supplier admission.
+
+## Evidence and controls
+
+`accounting.py` is standard-library Python and has seven input-sensitive rejection controls:
+
+- Delete the third child's extra whole idle.
+- Give the fresh suffix an unearned central Gaussian factor.
+- Delete a fresh stage from that suffix target.
+- Delete one rank-one query.
+- Delete one native invocation, invalidating cancellation.
+- Replace an ordinary Q² translation with a paid odd Gaussian exponent.
+- Exchange the retired/fresh field targets, preserving their histogram but violating their interface.
+
+Both actual sources are positive fixtures. Normal and optimized Python 3.12.14 executions passed with byte-identical JSON. Measured subprocess times were approximately 0.08 s and 0.21 s; maximum child RSS was below 35 MiB. Exact commands, status and timings are in `evidence/accounting-normal-run.json` and `evidence/accounting-optimized-run.json`; the deterministic scientific result is `RESULT.json` (the accounting subsection in this integrated package).
+
+From a checkout with the frozen directories staged under the indicated paths:
+
+```sh
+python -I -B accounting.py --predecessor PATH/gaussian-root-retirement --candidate PATH/retirement-rebalanced-query-tree
+python -I -O -B accounting.py --predecessor PATH/gaussian-root-retirement --candidate PATH/retirement-rebalanced-query-tree
+```
+
+## Exact dependencies and limits
+
+Runtime reads:
+
+- Predecessor: `CONTROLLER.json`, `WITNESS.json`. The native archive is read in memory from the candidate's exact retained `source.tar`, and its independently observed predecessor SHA256 is checked.
+- Candidate: `CONTROLLER.json`, `ROOT.json`, `WITNESS.json`, `source.tar`.
+- Nested old ROOT: `native-source.tar!research/p104-scaled-transport-fusion/source.tar!research/p104-third-anchor-clean-mixed-root/root-source.tar!research/mixed-projector-query-root/ROOT.json`.
+- Nested native word: `native-source.tar!research/p104-scaled-transport-fusion/WORD.json.gz`.
+
+No archive paths are extracted. Source-pin authentication should surround this checker in the integrated package. The native archive is checked against the independently observed predecessor hash; the other reported SHA256 values identify bytes actually read and need the integrated source pins to bind GitHub heads.
+
+The caller identity is a separate obligation: this checker assumes the physical meaning of the serialized endpoint rows, and checks accounting/interface consistency rather than proving those endpoints from gate composition. It does not replay native phase maps, verify the retirement prefix/dual-pair identity, prove arbitrary-input restoration, certify the numerical root, validate changed completion decay, implement a uniform compiler, price total ordinary work, or resolve any of the five framework assumptions. Those obligations must remain separately labelled even if this delta check passes.
+
+Credit Ali Romman for PR6551 and PR6545, the PR6398 mixed-root and PR6529 native-source lineage, and all inherited source notices. This independent accounting implementation and explanation were prepared for DLamdata with substantial OpenAI Codex assistance. Apache-2.0; no independent human or proof-kernel verification is claimed.
diff --git a/research/pr6551-retirement-audit/DOCUMENTATION-REVIEW.md b/research/pr6551-retirement-audit/DOCUMENTATION-REVIEW.md
@@ -0,0 +1,24 @@
+# Read-only documentation review
+
+Reviewed README.md, EXTERNAL-REVIEW.md, PRICING-SOURCE-MAP.md, ENDPOINT-REVIEW.md and pricing_binding.py against the supplied external audit text. No scientific verifier or artifact-regenerating command was run. The extended checker hash in ENDPOINT-REVIEW.md matches the current file (`3b15f6c17d2e7a941c97a29cd4269388cd07f824a4d8df16fc22a62c1515fcf8`). Publication receipts are intentionally outside this interim check.
+
+## Findings to resolve
+
+1. **Minor factual source-map correction — PRICING-SOURCE-MAP.md, paragraph beginning “The candidate's verify.py”.** It says `native_word`, `local`, `native`, and `L` all pass to both the predecessor invocation and the new price calculation. Exact candidate code passes `(native_word, old_caller, local, native, L)` to `source_candidate.run` at line 88, but only `(local, native, word, physical, L)` to `price.run` at line 136. The native word instead goes to `physical(..., native_word, L)` at line 131. Suggested replacement: “The candidate's `verify.py` `Extension.run` passes the same `local`, `native` and `L` objects to the predecessor invocation at line 88 and the new price calculation at line 136; the same `native_word` is supplied to the physical-map check at line 131.” This does not alter the cancellation conclusion.
+
+2. **Attribution clarity — README.md, Pins and independence final paragraph.** The sentence “The reviewer made the conceptual suffix argument's implicit invertibility premise explicit…” follows the Claude paragraph, so a reader can attribute that original internal-review finding to Claude. Earlier internal review supplied it. Change “The reviewer” to “The internal skeptical reviewer” (or name that role explicitly). Claude's six documented findings and later Codex implementations are otherwise cleanly distinguished.
+
+## Checks with no issue found
+
+- Claude Code / Claude Opus 5.5 is credited for the external audit and findings; GPT-6 Astra through Codex is credited for the lead/internal roles and follow-up implementation. No independent human or formal-kernel review is claimed.
+- External results are explicitly **reported** at `13e880a734cc03f2a52c82a412a9a0e6ac2b1f0c`; unpublished scripts/logs are not represented as locally replayed evidence. The later full-fiber checker is correctly separated from that external snapshot.
+- The external report's early contradictory progress wording (“caught” versus “accepts”, “passes” versus “past” the decay) is resolved using its final substantive findings: four suffix changes passed the old cut check, and the candidate ratio exceeds 31/32.
+- The additional local full-endpoint coverage and four suffix rejection controls are identified separately from the old cut check; redundant guards are not falsely described as independently mutation-covered.
+- Physical tariff validity remains inherited. The exact 232,960 ordinary allowance increment is presented as arithmetic under the pinned formula, not a proof of physical rates or supplier admission.
+- Local-manifest integrity is distinguished from external Git-object provenance, with an explicit trusted-publication/checker boundary. The AST matching helper calls its fragments copied source contracts, not independently derived physical semantics.
+- `research/20261103-open-obligation` is correctly the branch, and `research/pr6551-retirement-audit` the package directory. The external report's initial branch confusion was corrected and is not repeated.
+- The old 20-file/53-file external snapshot and current 22 retained sources are distinguished. Historical receipts are labelled; this review intentionally does not flag pending final publication receipts.
+
+No blocking scientific or scope-overstatement issue was found in these documents. Resolve the two small factual/attribution edits before publication, then let the lead's serial validation and final receipt process complete.
+
+Lead resolution: both wording findings were corrected before the scientific follow-up commit. The source-map line now distinguishes the shared pricing arguments from native_word passed to the physical checker; README names the internal skeptical reviewer. No scientific code or receipt changed.
diff --git a/research/pr6551-retirement-audit/ENDPOINT-REVIEW.md b/research/pr6551-retirement-audit/ENDPOINT-REVIEW.md
@@ -0,0 +1,58 @@
+# Independent full-endpoint extension after external review
+
+**Result:** both frozen callers pass independent full 35-by-35 endpoint verification in all four exact central fibers, with every stage variable formal. The previous retirement-cut and input-lifetime checks remain. The four requested changed-suffix controls all pass those previous checks but fail the new endpoint calculation with explicit coefficient counterexamples.
+
+This extends the review coverage; it does not change either caller, source claim, invoice, source pin or selected frontier. No predecessor constructor, source checker, lead checker or pricing routine is imported or executed. The new engine reuses only this reviewer's own Gaussian-rational, formal-stage vector interpreter.
+
+## Exact scope
+
+The retained inputs remain PR6545 `775722aa94a5e85b05d1d704d7dca10cbf016f0c` and PR6551 `11ca568cd1a16776ade02b4df09ea04a784cc674`, reviewed through contribution snapshot `13e880a734cc03f2a52c82a412a9a0e6ac2b1f0c`:
+
+- Old controller SHA256: `e5c873568b52aab82fc01670de3037b1ba4bedbcb15366f0e5fb38cd786714ed`.
+- New controller SHA256: `419566fe8d8e84a21a14b8efbdac83df5cf1892b0de55f4f256287e9031d74da`.
+- Extended checker SHA256: `3b15f6c17d2e7a941c97a29cd4269388cd07f824a4d8df16fc22a62c1515fcf8`.
+- Deterministic receipt SHA256: `ff7621c206e1a553c95e41db7a13cd70877aaf4646ccab2d9f603620adcfb5a9`.
+
+The standalone checker reports controller hashes but does not check their expected values itself. The integrated wrapper retains its source-integrity checks before launching this checker; independent source provenance requires the optional Git-object comparison.
+
+## Complete endpoint argument
+
+For each q in `1,-1,i,-i`, the checker starts with each of the 35 unit output rows and propagates it backward through every event (609 old, 611 new). Every one of the 35 resulting input coefficients is compared with its prescribed diagonal coefficient. This is 1,225 formal coefficient positions per fiber, 4,900 per caller, and 9,800 across the two actual callers. The coefficient representation is independent of the lead's complete transfer matrix: each row is processed separately after exact q specialization, with Gaussian rational coefficients and 19-component formal stage-exponent tuples. D0 is retained explicitly and required to be zero in target signatures; D1 through D18 are never numerically specialized.
+
+The existing Chinese remainder argument applies to the full matrices: over `A=Q(i)[D0,...,D18]`, the four factors of `q^4-1` are pairwise comaximal, so evaluation gives an injection `A[q]/(q^4-1) -> A^4`. Exact equality in all four fibers therefore establishes each complete endpoint identity in the claimed formal quotient ring. This is not random testing or a specialization of the stage operators. The written CRT argument is not a formal-kernel proof.
+
+The matrices still use the declared LOCAL, QUERY, IDLE, LAMBDA and ordinary scalar contracts. Physical native maps, all-width substitution, scalar-outside-tensor convention, ordinary work, numerical moments, compiler and outer obligations remain separate. Agreement with the primary replay is implementation corroboration under shared semantic and Python Fraction assumptions, not independent physical admission.
+
+## Independently identified changed region
+
+A maximal literal comparison gives exactly 358 common initial events and 67 common final events. The changed half-open slices are old `[358:542]` and new `[358:544]`. Changed diagonal roles are `20,21,22,23,24,25,26,27,34`. The 345-event retirement prefix is inside the unchanged prefix; this explains why the earlier cut-only checker alone could not detect the requested suffix errors. Identical trailing instructions do not imply identical incoming states, so the extension checks the complete end-to-end map.
+
+## Requested failure controls
+
+Each control first passes the existing cut and lifetime checks, then must raise the specifically typed `EndpointMismatch`. An input-validation or cut-marker failure is not accepted as evidence of suffix coverage. The receipt records the first exact mismatch, including its full formal polynomial, expected polynomial, q fiber and matrix coordinates.
+
+| Invalid change | First mismatch at q=1 | Detection |
+|---|---|---|
+| Flip event363 `ADD 1 22 1` to coefficient -1 | output23, input20 | Actual coefficient is `2 D10 D11`, expected zero. |
+| Remove D11 from event358's five-stage IDLE | output0, input0 | Actual stage monomial differs from the full target. |
+| Swap event371 LOCAL source and destination | output0, input0 | Actual diagonal coefficient differs from the required one. |
+| Remove D11 from fresh coordinate34's target | output34, input34 | Literal caller retains D11; declared target omits it. |
+
+These four controls preserve the first345 events and pass the original retirement-cut checks. Their endpoint failures are thus direct evidence for the added scientific coverage. The controls stop at the first disagreement; they do not claim a replay of every later mismatch.
+
+Both actual source callers are positive fixtures. A second new positive fixture inserts an inverse half-ADD pair on arbitrary fresh input34 into the changed suffix after event371; all 4,900 endpoint coefficient positions still pass. This tests a semantically valid suffix edit instead of simply rejecting every deviation from the pinned word. The previous canceling-ADD prefix fixture and four previous cut/lifetime rejection controls remain. An additional artificial pivot-column failure was unnecessary: the retained column check remains, and all complete coefficient positions now receive independent coverage.
+
+## Actual commands and results
+
+From the intended checkout after environment activation:
+
+```sh
+python -I -B /workspace/work/research-assignment/claude-followup/fibers/cut_four_fibers.py /workspace/work/research-assignment/checkout/research/pr6551-retirement-audit/source/pr6545/CONTROLLER.json /workspace/work/research-assignment/checkout/research/pr6551-retirement-audit/source/pr6551/CONTROLLER.json
+python -I -O -B /workspace/work/research-assignment/claude-followup/fibers/cut_four_fibers.py /workspace/work/research-assignment/checkout/research/pr6551-retirement-audit/source/pr6545/CONTROLLER.json /workspace/work/research-assignment/checkout/research/pr6551-retirement-audit/source/pr6551/CONTROLLER.json
+```
+
+Both exit0 with `PASS_FOUR_FIBER_RETIREMENT_CUT_AND_ENDPOINTS`; normal and optimized receipts are byte-identical. CPython 3.12.14, standard library only. Normal run: 1.099399 seconds, 11904 KiB child high-water RSS. Optimized run: 1.157566 seconds, 14788 KiB. The measurement wrappers used a 120-second timeout and Python `time.monotonic`/`resource.getrusage`. Full executed interpreter paths, arguments, exit status, stderr and measurements are in [fiber-normal-run.json](evidence/followup/fiber-normal-run.json) and [fiber-optimized-run.json](evidence/followup/fiber-optimized-run.json). No scientific run failed.
+
+The extension was developed and run only in the assigned separate output directory. No project file, Git ref, public message or PR was changed by this agent. The lead integrated the new script/receipt; [FOLLOWUP-VALIDATION.md](FOLLOWUP-VALIDATION.md) records the serial registered package, export and publication-byte checks. The previous reviewed source/caller snapshot is unchanged, while this expanded checker snapshot requires its own evidence.
+
+Prepared for DLamdata with substantial GPT-6 Astra assistance through OpenAI Codex in the independent review role; Apache-2.0. Ali Romman and the retained predecessor contributors keep construction, retirement and native-interface credit. This is AI review, not human or maintainer acceptance.