Corroborating audit of PR6030: fresh full dense replay, itinerary delta, invoice and crossing (no new κ)

research/p104-6030-independent-audit/NEW-MAPS.json

@@ -0,0 +1 @@
+{"schema":"p104-6030-new-dense-maps/1","maps":[[["helper",3,189,0],8,"065ffe910bec1d3b661770aeb45d3f331fcf85f2f237b562288bab13d77d82ea"],[["helper",76,394,0],2,"0ebdac8dae67b1e89622c4c0ffde56835a6ccef46d4b64e8e3d947f70dab4c11"],[["helper",749,127,0],6,"7a5100f1e66d3fd47e1eb9ee69a69dde29ba76b95b1173440285636dfce10c14"],[["move",3,189,0],8,"065ffe910bec1d3b661770aeb45d3f331fcf85f2f237b562288bab13d77d82ea"],[["move",3,714,0],4,"07c471517840c82c5b5778f77cae7765829c6927aa8e9d2ed059938d944885aa"],[["move",595,637,0],2,"2d540100a251ff983e5ccf9d96f8215a389c70759923a0813673efeb35bc126b"],[["move",714,127,0],5,"0b207429a4ef9b8992645a564cf9e8ed433ac6acfda523be0acda7e6e40e0c50"]]}

research/p104-6030-independent-audit/README.md

@@ -0,0 +1,95 @@
+# Independent replay, delta, invoice, native and funding audit of PR6030
+
+**Result (scoped corroborating audit, no new κ).** An independent check of
+[PR #6030](https://github.com/CrocSwap/integer-mult-bounds/pull/6030) (DreamingOfClouds, helper 470's frame-593 rung
+removal, conditional κ 0.003125804) finds no error. On canonical main, an own exact replay, an exact structural
+PR #5907 → PR #6030 delta, an own invoice recount, a fresh dense all-column native check with absolute phase of
+**every** operator in the PR6030 inventory and an own decimal funding engine all reproduce the author's claims.
+That includes the raw crossing (0.00312581155791, 0.00312581155792). The per-LOCAL margin change is H1 + H3 − H4,
+the same as PR #5907's gain over PR #5827.
+
+**Relation to the admission audit.** This work started from `a2b4b039`, where the maintainer
+[repair review](../../docs/research/pr-intake-evidence/pr6030-repair/REVIEW.md) left an independent check of the
+changed word open. While it ran, the maintainer admitted PR6030 at `721c0382`, after an independent application
+audit (`docs/research/pr-intake-evidence/pr6030-audit/PROOF-REVIEW.md`). That audit also uses Matarener's
+interpreter and dense inventory code. So this package is a second execution of largely the same checks, not new
+coverage, and it changes no decision. It adds three things. First, `--full` recomputes the dense digest of all 3,163
+operators. The admission review checks only the 7 new tasks densely and says it "is not a fresh independent dense
+replay of every unchanged predecessor map". Second, an itinerary-level delta: exactly six register itineraries
+change, and every other MOVE is unchanged as a multiset. Third, five tamper controls on the PR6030 word.
+
+Base: canonical main `721c03822c421b7f7fcbb2598993b947c4d88419` (the work ran at `a2b4b039`; no input changed between the two). This contains PR #6030's package
+(`research/p104-helper470-twin-rung-removal`, construction head `772dd10a`, with the maintainer portability repair
+`746ee5e1`) and PR #5907's (`research/p104-helper468-rung-removal`). Live inputs are pinned in
+[SOURCES.json](SOURCES.json).
+
+## What is checked
+
+The replay, invoice, dense-comparison and funding functions are copied unchanged from Matarener's
+[PR5827/PR5907 audit](../p104-unread-rung-independent-audit/README.md) (PR #5960). Its digests for 3,156 of the 3,163 PR6030
+native operators are reused. Nothing is imported from the PR6030 package except its data files and, in layer 0
+only, its `construct.py`. The PR #5827 verifier chain, which the author ran, is not used anywhere.
+
+| Layer | Check | Result |
+|---|---|---|
+| 0. Sources | PR6030's `inputs/pr5907-WORD.json.gz` and `inputs/pr5827-source.tar` against canonical main's PR5907 word and PR5907's pinned archive; PR6030 `construct.py` applied to canonical PR5907 word with `EDIT.json.gz` | pinned word and archive byte-identical; own serialization of the regenerated word equals the decompressed shipped payload byte for byte |
+| 1. Own replay | PR5907 and PR6030 words: strict span nesting of every MOVE, common chart of every ADD/COPY/SCALE, dyadic coefficients, owned INIT/CLEAR lifetimes, all 799 endpoint rows (identity on public registers plus X into Y for the 104 live ports), helper initial span strictly inside final span | PASS on both: 770 public, 29 owned fields, 31 lifetimes, peak 13; paid and helper histograms equal each author's `local` block |
+| | PR5907 → PR6030 delta (structure) | all fields except `program`, `initial`, `final` unchanged; exactly the three declared boundary changes (initial 713: 3→76, 736: 424→749; final 115: 105→189); exactly the six declared itineraries change (registers 115, 127, 470, 704, 713, 736, as listed in the PR6030 README); INIT, CLEAR, COPY and every other MOVE unchanged as multisets; paid delta −2H1 + H2 − 2H3 + H4 + H6 − H7 + H8, helper delta −H1 + H2 − H3 + H6 − H7 + H8 (helper ranks 115: 1→8, 713: 3→2, 736: 7→6); 22,803 events ≤ 26,000. The rewritten ADD and SCALE events (ADD 21,706 → 18,404, SCALE 753 → 861) are not compared with a declared edit; they are checked only through their exact effect in the replay above, which is what the endpoint identity needs |
+| | tamper controls | flipped ADD sign (helper 268), dropped ADD (helper 470), non-dyadic SCALE, restored final chart of 115 and a reinstated frame-593 MOVE for 470 are each rejected |
+| 2. Invoice | own recount from the 51 physical caller events (18 LOCAL, 15 QUERY_SHEAR, 15 IDLE, diagonal terms) and separately from the closed formula in PR5827 PROOF.md | equal to each author `complete_cost`: 66,852 / 13,444 (PR5907) and 66,834 / 13,444 (PR6030); rank-mass surplus 760 |
+| 3. Dense native | PR5548's credited exact Gaussian-dyadic interpreter compares the lowered recipe with the literal frame circuit on all 512 columns: exact equality, norm, support size 2^rank | PR6030 inventory: 3,163 operators (2,886 distinct MOVE maps, 69 distinct helper pairs, 208 main adapters). The 2,922 distinct signed maps and 1,124 helper normalizers equal the author's `native` block. The 7 maps new in PR6030 are computed here (helper 3→189, 76→394, 749→127; MOVE 3→189, 3→714, 595→637, 714→127) and pinned in [NEW-MAPS.json](NEW-MAPS.json). Every other map equals the predecessor's FULL-CHECK digest. A phase-only fault and a flipped child sign on the new map 3→189 are rejected |
+| 4. Own funding | 80-digit decimal ln/exp with 10⁻⁷⁰ relative padding, bisection on a 10⁻¹⁴ grid | raw crossings of PR5907 and PR6030 equal the author intervals; tested saving .0031258090 positive; tail ratios α < 1 − 10⁻¹², γ, δ < 125/128; margin derivative negative on [0, .0034], so the crossing is unique there; claim gain 261/500000000 |
+
+The gain over PR5907 per LOCAL call is H1 + H3 − H4 in the moment margin, with H_r = r^t and t = 1 − κ. It is
+positive for every t in (0, 1), because (1/4)^t + (3/4)^t decreases in t and equals 1 at t = 1.
+
+## Reproduce
+
+Python 3.12, standard library only, from the repository root:
+
+```sh
+python3.12 -B research/p104-6030-independent-audit/verify.py
+```
+
+The default run is serial and takes about one minute (61 s on Python 3.12, Apple M-series; it also passes on 3.10, 3.11, 3.13 and 3.14). It checks layers 0, 1, 2 and 4 completely. For layer 3 it
+computes all 7 new maps and a fixed subset of the old ones: every helper pair, every 8th main adapter and every
+64th MOVE map, each matched against the predecessor's FULL-CHECK.json. The complete PR6030 inventory runs with
+
+```sh
+python3.12 -B research/p104-6030-independent-audit/verify.py --full --jobs 8
+```
+
+in about 4 minutes on eight cores (4 min 14 s, Python 3.12). It recomputes all 3,163 digests. The marker is `PASS_P104_6030_INDEPENDENT_AUDIT`.
+
+## Scope and limits
+
+This audits the finite 9-qubit local word and its complete invoice. The five inherited framework assumptions remain:
+native_all_columns for the uniform word, owned_erasure, compiler, precision_tape and analytic_recovery. The physical
+caller (`research/p104-productive-controller`) is recounted, not re-derived. That changing a helper's initial or
+final chart (here 713, 736 and 115) is a free boundary choice is inherited from the P104 line; PR5907 used the same
+freedom. The compiler threshold and the 54 strict and 4 functional outer conditions are checked only by the
+author's chain. The maintainer ran that chain in full on macOS and hosted Linux; it is not rerun here. The layer 1
+to 4 code is shared with the predecessor audit, so it is not independent of that audit, only of the PR6030 author.
+This is an agent check, not human or proof-kernel acceptance. It changes no decision: the selected record was already PR6030 at `721c0382`.
+
+## Open next questions
+
+- Can the outer conditions and the compiler threshold be recomputed by code that does not descend from PR5827's
+  `pricing.py`? This package and its predecessor check them only through the authors' chain. The admission review
+  records that all 54 strict and 4 functional predicates pass, but does not say which code computes them. First step: write each named margin
+  in `complete_cost.outer` as an exact rational at κ = 781451/250000000 from its source document. Blocked on finding
+  the formula source for each margin.
+- Does the uniform compiler preserve the checked finite native maps at every recursion level? This is blocked by
+  the inherited native_all_columns premise.
+
+## Credit
+
+DreamingOfClouds (PR6030 and PR5907 edits and packaging). Ali Romman (PR5827 word, verifier and lineage, with
+OpenAI Codex assistance), Mason Hale (PR5811) and the lineage back to Pierre-Loic Doulcet (PR5565). Matarener
+(PR5960, the PR5827/PR5907 independent audit, whose verifier functions and FULL-CHECK digests this package reuses).
+Chris R / Clear1nsight (PR5548 dense interpreter, imported unchanged). Douglas Colkitt with OpenAI Codex
+(the PR6030 portability repair and primary replay). The P104 native, Clifford, controller, compiler and analytic
+contributors are credited in the cited packages.
+
+Prepared for JosieElliston with Claude Opus 5.5 (Anthropic), via Claude Code, in an unattended donated session.
+Apache-2.0.