diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/DELIVERY_MANIFEST.json b/audits/exact_sample_shrinking_bin/2026-10-24/DELIVERY_MANIFEST.json new file mode 100644 index 000000000..b32270a2c --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/DELIVERY_MANIFEST.json @@ -0,0 +1 @@ +{"files":[{"filename":"HANDOFF.json","sha256":"3781c64e791a7206032b8a16d4a7a5668037517c3a61ec58c23df97af431b40d","size":3736},{"filename":"README.md","sha256":"c62694552ab35bb6e375bcbe6690bc5aadde2b2cab0027869903c4899f651c4d","size":5573},{"filename":"controls/esb_check.py","sha256":"01127db3ad48fb7a30f04afff99daad9a5f1e82c88cf0e4a9ce7430b646598e7","size":9937},{"filename":"controls/esb_check.stdout.txt","sha256":"cf1ff68d9bd1c1759ffb1e8629e73b292ad268bca4f33f0ed02720796d4f8835","size":536},{"filename":"sources/comment-5976299711.body.txt","sha256":"819bc7c7fd637883ca74626b01d777d40233f6b710b6f26b5d370648e09a945d","size":6104},{"filename":"sources/comment-5976527971.body.txt","sha256":"ad6f20ab98ce288c2ab54f88653df908015a599f30a329f3c1f12eff93e7f16f","size":9103},{"filename":"sources/comment-5976979978.body.txt","sha256":"7f2eaffe5195466b94366959fc8f6c40bcebe76d5c50e2fe453f12f5bdeb1c0e","size":7708},{"filename":"sources/comment-5976980568.body.txt","sha256":"b1c697378866b8d5b5c370593c1be128bf819fd91ab365837ee7f27de10ad2a1","size":11377}],"handoff":"HANDOFF.json","version":1} \ No newline at end of file diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/HANDOFF.json b/audits/exact_sample_shrinking_bin/2026-10-24/HANDOFF.json new file mode 100644 index 000000000..aece78985 --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/HANDOFF.json @@ -0,0 +1,77 @@ +{ + "payload": { + "protocol_version": "R17-immutable-handoff-v1", + "checkpoint_utc": "2026-10-24T20:08:29Z", + "task_id": "ESB-COMMENT-SOURCE-CUSTODY-20261024", + "claim_id": "https://github.com/d6g8k5htny-coder/main/pull/527#issuecomment-6388918019", + "stable_target_key": "main:audits/exact_sample_shrinking_bin/2026-10-24", + "source_identities": [ + { + "path": "sources/comment-5976299711.body.txt", + "source_url": "https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976299711", + "source_created_at": "2026-10-04T03:51:07Z", + "source_updated_at": "2026-10-04T03:51:54Z", + "bytes": 6104, + "sha256": "819bc7c7fd637883ca74626b01d777d40233f6b710b6f26b5d370648e09a945d", + "terminal_lf_count": 0 + }, + { + "path": "sources/comment-5976527971.body.txt", + "source_url": "https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976527971", + "source_created_at": "2026-10-04T04:21:12Z", + "source_updated_at": "2026-10-04T04:21:12Z", + "bytes": 9103, + "sha256": "ad6f20ab98ce288c2ab54f88653df908015a599f30a329f3c1f12eff93e7f16f", + "terminal_lf_count": 1 + }, + { + "path": "sources/comment-5976979978.body.txt", + "source_url": "https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976979978", + "source_created_at": "2026-10-04T05:36:04Z", + "source_updated_at": "2026-10-04T05:36:20Z", + "bytes": 7708, + "sha256": "7f2eaffe5195466b94366959fc8f6c40bcebe76d5c50e2fe453f12f5bdeb1c0e", + "terminal_lf_count": 0 + }, + { + "path": "sources/comment-5976980568.body.txt", + "source_url": "https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976980568", + "source_created_at": "2026-10-04T05:36:10Z", + "source_updated_at": "2026-10-04T05:36:10Z", + "bytes": 11377, + "sha256": "b1c697378866b8d5b5c370593c1be128bf819fd91ab365837ee7f27de10ad2a1", + "terminal_lf_count": 0 + }, + { + "path": "controls/esb_check.py", + "source_url": "https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976980568", + "extraction": "first python fence capture excluding LF before closing fence, then append exactly one LF", + "bytes": 9937, + "sha256": "01127db3ad48fb7a30f04afff99daad9a5f1e82c88cf0e4a9ce7430b646598e7", + "execution": "NOT_RUN_BY_THIS_STAGE" + }, + { + "path": "controls/esb_check.stdout.txt", + "source_url": "https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976980568", + "extraction": "first text fence capture excluding LF before closing fence, then append exactly one LF", + "bytes": 536, + "sha256": "cf1ff68d9bd1c1759ffb1e8629e73b292ad268bca4f33f0ed02720796d4f8835", + "execution": "NOT_RUN_BY_THIS_STAGE" + } + ], + "output_head": { + "kind": "prepublication-source-snapshot", + "parent_commit": "a38707383d406e0751481f3d0152e71b527c0c9a", + "source_set_sha256": "d5b7aaf32224fdb1c63847febf6d4a9f632ce1fd7e48367238ab43d24e150268" + }, + "verified_boundary": { + "scope": "exact decoded public-body custody and declared block extraction only", + "program_execution": "NOT_RUN", + "historical_stdout": "copied original record, not a new result", + "scientific_effect": "NONE", + "organizational_independence_credit": 0 + }, + "next_action": "Separate nonauthor source-custody review and protected integration. Original comments remain canonical; no mathematical acceptance or execution authority supplied." + }, + "payload_sha256": "5fb845854cb8c57c2a9f77e983254ae6ec4b3d89524f6bde279625ad6a9e2486" +} diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/README.md b/audits/exact_sample_shrinking_bin/2026-10-24/README.md new file mode 100644 index 000000000..d085ee38b --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/README.md @@ -0,0 +1,85 @@ +# Exact-sample comment sources and historical reviewer controls + +This is a byte-preservation package for existing public research records. The +original GitHub comments remain canonical. Their copies and the extracted +control artifacts here establish source custody, not a new proof, review, +execution, scientific acceptance or certification of stored pilot samples. + +## Preserved objects + +Each `sources/*.body.txt` is exactly the decoded GitHub REST `body` encoded as +UTF-8: no surrounding JSON, HTML, newline addition or Unicode normalization. + +| Comment | Recorded role | Bytes | SHA-256 | +|---|---|---:|---| +| [5976299711](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976299711) | Original OpenAI/Codex author-side proof | 6104 | `819bc7c7fd637883ca74626b01d777d40233f6b710b6f26b5d370648e09a945d` | +| [5976527971](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976527971) | Existing OpenAI/Codex scoped review | 9103 | `ad6f20ab98ce288c2ab54f88653df908015a599f30a329f3c1f12eff93e7f16f` | +| [5976979978](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976979978) | Existing Anthropic Claude scoped review | 7708 | `7f2eaffe5195466b94366959fc8f6c40bcebe76d5c50e2fe453f12f5bdeb1c0e` | +| [5976980568](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976980568) | Claude reviewer's published control program and recorded output | 11377 | `b1c697378866b8d5b5c370593c1be128bf819fd91ab365837ee7f27de10ad2a1` | + +The source records supply the authorship and review exposure above. Shared +account publication authenticates neither distinct operators nor organizational +independence. Both historical technical reviews remain scoped to their consumed +interfaces; the Claude review discloses reading the earlier OpenAI review. +Its displayed ACCEPT/PASS label is a technical verdict, not scientific +acceptance. This preservation stage supplies no new mathematical verdict. + +`controls/esb_check.py` is the **original historical reviewer program**: +9937 bytes, SHA-256 +`01127db3ad48fb7a30f04afff99daad9a5f1e82c88cf0e4a9ce7430b646598e7`. +`controls/esb_check.stdout.txt` is the **original recorded stdout**, not output +of a run performed for this package: 536 bytes, SHA-256 +`cf1ff68d9bd1c1759ffb1e8629e73b292ad268bca4f33f0ed02720796d4f8835`. +Their contents, including the program's Anthropic attribution and the output's +historical ALL PASS line, are preserved unchanged. + +For each extraction, use the first fenced block with the declared language in +comment 5976980568, capture the bytes between the opening fence's LF and the LF +immediately before the closing fence, then append exactly one LF. The languages +are `python` for the program and `text` for stdout. This rule reproduces both +original published hashes. No source program or mutant was executed here. + +## Three different consumer relationships + +The [existing reading record](../../../experiments/periodic_h0/EXACT_SAMPLE_SHRINKING_BIN.md) +routes readers to the complete comment proof and its source-bound reviews. +The comment proof is consumed analytically by +[`spectral_shrinking_bins/PROOF.md`, section 4](../../../experiments/periodic_h0/spectral_shrinking_bins/PROOF.md) +under that extension's own source map. The historical reviewer program checks +finite/algebraic/rate-bookkeeping controls described in its header; it is not an +implementation or verification of the complete stochastic asymptotic theorem. +The [publication-fidelity test](../../../tests/test_exact_sample_shrinking_bin.py) +checks the repository reading route and recorded identities, not the proof's +mathematical validity. These are separate relationships. + +Existing hypergraph row 307 already represents the periodic-H0 container. Its +generic `controls.py`/`count_confidence.py` membership does not certify that +those programs implement this lemma. This package changes no row, graph edge, +scientific register, acceptance flag, parent-source qualification or pilot +result. It does not supply spectral/FFT/roundoff/library error control, numerical +constants, a held-out experiment, continuum moment bounds or parent-theorem +acceptance. The original comments retain their full hypotheses and exclusions. + +## Custody and reproduction + +Source retrieval was checked at Main +`a38707383d406e0751481f3d0152e71b527c0c9a`. No exact current Git carrier was +found in the checked Main/Math trees; complete history-wide absence is not +claimed. The package is a new Git carrier, not regenerated historical work. +All inputs were already public project records; no private storage was read or +published for this package. + +`HANDOFF.json` uses the existing R17 delivery envelope with this task's public +GitHub pickup as its claim ID, not an invented native Work Events record. Its +output identity describes the pre-publication source snapshot; the actual Git +commit mapping belongs in the PR's delivery receipt. `DELIVERY_MANIFEST.json` +is generated by the existing `tools/research_delivery.py`, with no self-hash. +The private deterministic ZIP built from the explicit eight-file input list was +verified without extracting or executing payloads. The same generated manifest +is retained here beside its eight inputs. Byte/manifest verification establishes +custody only. This package grants no execution or integration authority. + +Actual preserver: OpenAI/Codex `/root`, under the existing owner mandate and +delegation. Historical authorship is unchanged. Separate nonauthor custody +review and normal protected integration remain necessary. credit 0; scientific +effect NONE; OBL OPEN; NOT A RESULT. diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/controls/esb_check.py b/audits/exact_sample_shrinking_bin/2026-10-24/controls/esb_check.py new file mode 100644 index 000000000..9244ccf6b --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/controls/esb_check.py @@ -0,0 +1,185 @@ +#!/usr/bin/env python3 +"""Reviewer controls for the exact-sample shrinking-bin lemma (main#229 5976299711). Standard library only; exact +rationals; deterministic seeds. Anthropic Claude, cross-provider nonauthor review. Scientific effect NONE. + +Groups: + S the sharp-bin injection (1): random finite diagrams P, explicit epsilon-matchings to Q (pairs moved by at most + epsilon per endpoint, extra short bars matched to the diagonal), random half-open bins with a > d = 2 epsilon, + many exact endpoint coincidences; check |N_Q(I) - N_P(I)| <= N_P(S_d(I)), S_d(I) = [a-d,a+d) u [b-d,b+d). + T tightness and necessity: equality cases exist, and with a <= d the bound can fail (a single bar of lifetime d). + G the triangle geometry behind B_h (APPROXIMATION): sum_i lambda_i |v_i - x|^2 = x1(h-x1) + x2(h-x2) <= h^2/2 on both + triangles of both diagonals (exact rationals on a grid of points). + D |det M| <= ||M||_op^2 for 2 x 2 symmetric M (exact: compare det^2 with the squared largest eigenvalue squared). + M the bin mass: integral of l^(-1/3) over [lam tau, mu tau) = (3/2)(mu^(2/3) - lam^(2/3)) tau^(2/3) (Simpson check), and + the strip bound: total strip length <= 4 d, strip lifetimes >= lam tau - d. + R rate bookkeeping of section 5 for schedules tau_h = h^2 sqrt(log(1/h)) w_h with w_h -> infinity: the two good-event + ratios d_h/tau_h and d_h/tau_h^(2/3) tend to 0, and the bad-event ratio (h^-2 + 1) h^(c A^2) / tau_h^(2/3) tends to 0 + exactly when c A^2 > 10/3 (checked on both sides of the threshold). +Mutants (--mutant X) must make the run exit 1: HALF_STRIPS (strips cut to [a-d,a) u [b-d,b)), A_GE_D (allow a = d), +WRONG_GEOM (h^2/4 instead of h^2/2), DET_POWER1 (the false bound |det M| <= ||M||_op), RATE_103 (threshold 3 instead of +10/3). Unknown labels exit 2.""" +import math, random, sys +from fractions import Fraction as F + +MUT = None +if len(sys.argv) == 3 and sys.argv[1] == '--mutant': + MUT = sys.argv[2] + if MUT not in ('HALF_STRIPS', 'A_GE_D', 'WRONG_GEOM', 'DET_POWER1', 'RATE_103'): + print('unknown mutant'); sys.exit(2) +elif len(sys.argv) != 1: + print('usage: esb_check.py [--mutant LABEL]'); sys.exit(2) + +fails = [] +def check(cond, msg): + if not cond: + fails.append(msg) + +def in_bin(l, a, b): + return a <= l < b + +def strips(a, b, d): + if MUT == 'HALF_STRIPS': + return [(a - d, a), (b - d, b)] + return [(a - d, a + d), (b - d, b + d)] + +def count(ls, lo, hi): + return sum(1 for l in ls if lo <= l < hi) + +def count_union(ls, ivs): + return sum(1 for l in ls if any(lo <= l < hi for lo, hi in ivs)) + +# ---------------------------------------------------------------- S: the sharp-bin injection +rng = random.Random(20261004) +nS = 0 +for trial in range(4000): + den = rng.choice([4, 8, 12, 24]) + eps = F(rng.randint(1, 6), den) + d = 2 * eps + P = [] + for _ in range(rng.randint(0, 12)): + birth = F(rng.randint(0, 200), den) + life = F(rng.randint(1, 80), den) + P.append((birth, birth - life)) # superlevel bar: birth > death + # matched images: each endpoint moves by at most eps (exact multiples of 1/den to hit endpoints) + Q, pairs = [], [] + for (bP, sP) in P: + while True: + db = F(rng.randint(-int(eps * den), int(eps * den)), den) + ds = F(rng.randint(-int(eps * den), int(eps * den)), den) + bQ, sQ = bP + db, sP + ds + if bQ - sQ > 0: + break + # a matched pair may also be sent to the diagonal only if both are short: keep it matched here + Q.append((bQ, sQ)); pairs.append(((bP, sP), (bQ, sQ))) + # extra Q bars matched to the diagonal: lifetime <= d + for _ in range(rng.randint(0, 4)): + bQ = F(rng.randint(0, 200), den) + Q.append((bQ, bQ - F(rng.randint(1, int(d * den)), den))) + # P bars matched to the diagonal would have lifetime <= d: drop matched P pairs that are short with prob 1/3 + lP = [b - s for b, s in P] + lQ = [b - s for b, s in Q] + for _ in range(6): + a = F(rng.randint(1, 60), den) + b = a + F(rng.randint(1, 60), den) + cond_a = (a >= d) if MUT == 'A_GE_D' else (a > d) + if not cond_a: + continue + lhs = abs(count(lQ, a, b) - count(lP, a, b)) + rhs = count_union(lP, strips(a, b, d)) + nS += 1 + check(lhs <= rhs, 'S: injection fails at trial %d (a=%s b=%s d=%s)' % (trial, a, b, d)) + +# ---------------------------------------------------------------- T: necessity of a > d, and an equality case +d = F(1, 2) +P, Q = [d], [] # one bar of lifetime exactly d, matched to the diagonal +a, b = d, 2 * d +lhs = abs(count(Q, a, b) - count(P, a, b)) +rhs = count_union(P, strips(a, b, d)) +# with a = d the bound happens to hold here (the P bar lies in the lower strip); the failing case is the reversed roles +P2, Q2 = [], [d] +lhs2 = abs(count(Q2, a, b) - count(P2, a, b)) +rhs2 = count_union(P2, strips(a, b, d)) +check(lhs2 == 1 and rhs2 == 0, 'T: the a <= d counterexample (continuum empty, approximation bar of lifetime d) must violate (1)') +if MUT == 'A_GE_D': + check(lhs2 <= rhs2, 'T: mutant allows a = d') +# equality: a matched pair straddling a +eP, eQ = [F(1)], [F(1) - F(1, 4)] +check(abs(count(eQ, 1, 2) - count(eP, 1, 2)) == count_union(eP, strips(F(1), F(2), F(1, 2))) == 1, 'T: equality case') + +# ---------------------------------------------------------------- G: triangle geometry of APPROXIMATION's B_h +h = F(1) +gmax = F(0) +for i in range(0, 25): + for j in range(0, 25): + x1, x2 = F(i, 24), F(j, 24) + for diag in ('AD', 'BC'): + if diag == 'AD': # diagonal (0,0)-(1,1): triangles below/above + tri = [(0, 0), (1, 0), (1, 1)] if x2 <= x1 else [(0, 0), (0, 1), (1, 1)] + else: # diagonal (1,0)-(0,1) + tri = [(0, 0), (1, 0), (0, 1)] if x1 + x2 <= 1 else [(1, 0), (0, 1), (1, 1)] + (a1, a2), (b1, b2), (c1, c2) = [(F(p), F(q)) for p, q in tri] + det = (b1 - a1) * (c2 - a2) - (c1 - a1) * (b2 - a2) + l2 = ((x1 - a1) * (c2 - a2) - (c1 - a1) * (x2 - a2)) / det + l3 = ((b1 - a1) * (x2 - a2) - (x1 - a1) * (b2 - a2)) / det + l1 = 1 - l2 - l3 + s = sum(l * ((p1 - x1) ** 2 + (p2 - x2) ** 2) for l, (p1, p2) in zip((l1, l2, l3), ((a1, a2), (b1, b2), (c1, c2)))) + check(s == x1 * (h - x1) + x2 * (h - x2), 'G: barycentric second moment identity') + gmax = max(gmax, s) +bound = h * h / 4 if MUT == 'WRONG_GEOM' else h * h / 2 +check(gmax <= bound, 'G: sum lambda |v - x|^2 exceeds the bound') + +# ---------------------------------------------------------------- D: |det M| <= ||M||_op^2 in d = 2 +rng2 = random.Random(7) +for _ in range(3000): + p, q, r = (F(rng2.randint(-50, 50), rng2.randint(1, 9)) for _ in range(3)) + det = p * r - q * q + # largest |eigenvalue|^2 = ((|tr| + sqrt(tr^2 - 4 det))/2)^2 >= |det|; exact test: |det| <= lam_max^2 iff ... + tr = p + r + disc = tr * tr - 4 * det # >= 0 for symmetric matrices + lam_max = (abs(float(tr)) + math.sqrt(float(disc))) / 2 + if MUT == 'DET_POWER1': + check(abs(float(det)) <= lam_max * (1 + 1e-12), 'D: mutant claims |det| <= ||M||_op') + check(abs(float(det)) <= lam_max ** 2 * (1 + 1e-12), 'D: |det| > ||M||^2') + +# ---------------------------------------------------------------- M: bin mass and strips +for lam, mu, tau in ((0.5, 2.0, 1e-3), (1.0, 3.0, 1e-5), (0.25, 0.75, 1e-2)): + a, b = lam * tau, mu * tau + n = 20000; hh = (b - a) / n + simpson = sum((1 if k in (0, n) else (4 if k % 2 else 2)) * (a + k * hh) ** (-1 / 3) for k in range(n + 1)) * hh / 3 + closed = 1.5 * (mu ** (2 / 3) - lam ** (2 / 3)) * tau ** (2 / 3) + check(abs(simpson / closed - 1) < 1e-9, 'M: bin mass') + dd = tau * lam / 4 + S = strips(a, b, dd) if MUT != 'HALF_STRIPS' else [(a - dd, a + dd), (b - dd, b + dd)] + check(sum(hi - lo for lo, hi in S) <= 4 * dd * (1 + 1e-15) and min(lo for lo, _ in S) >= a - dd, 'M: strips') + +# ---------------------------------------------------------------- R: rate bookkeeping of section 5 +thr = F(3) if MUT == 'RATE_103' else F(10, 3) +def bad_ratio_exponent(cA2): + # (h^-2) h^(cA2) / tau^(2/3) with tau = h^2 sqrt(log(1/h)) w, w -> infinity: the power of h is cA2 - 2 - 4/3 + return F(cA2) - 2 - F(4, 3) +for cA2 in (F(3), F(10, 3), F(7, 2), F(4)): + e = bad_ratio_exponent(cA2) + vanishes = e > 0 # at e = 0 the log factor log^(-1/3) still vanishes, but only slowly + check((cA2 > thr) == vanishes, 'R: threshold c A^2 > 10/3') +# numerical sanity of the three ratios along tau_h = h^2 sqrt(log(1/h)) log(log(1/h)) (w_h = log log 1/h), A = 3, c = 1/2 +prev = None +for k in range(4, 40, 4): + hv = 10.0 ** (-k / 4) + L = math.log(1 / hv) + tau = hv * hv * math.sqrt(L) * math.log(L) + dh = 1.5 * hv * hv * math.sqrt(L) + r1, r2 = dh / tau, dh / tau ** (2 / 3) + r3 = (hv ** -2 + 1) * math.exp(-0.5 * 9 * L) / tau ** (2 / 3) + if prev: + check(r1 < prev[0] and r2 < prev[1] and r3 < prev[2], 'R: ratios must decrease along the schedule') + prev = (r1, r2, r3) + +print('S: %d bin tests on random epsilon-matchings, injection (1) holds' % nS) +print('T: the strict condition a > d is necessary (a single approximation bar of lifetime d); an equality case is attained') +print('G: barycentric second moment = x1(h-x1)+x2(h-x2) <= h^2/2 on all four triangles (625 points each)') +print('D: |det M| <= ||M||_op^2 for 3000 random symmetric 2 x 2 matrices') +print('M: bin mass (3/2)(mu^(2/3)-lam^(2/3))tau^(2/3) to 1e-9; strip length <= 4d') +print('R: bad-event exponent c A^2 - 10/3; good-event ratios -> 0 along tau_h = h^2 sqrt(log 1/h) log log(1/h)') +if fails: + print('FAILURES:', len(fails)); print('\n'.join(sorted(set(fails))[:10])); sys.exit(1) +print('ALL PASS') diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/controls/esb_check.stdout.txt b/audits/exact_sample_shrinking_bin/2026-10-24/controls/esb_check.stdout.txt new file mode 100644 index 000000000..da516097f --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/controls/esb_check.stdout.txt @@ -0,0 +1,7 @@ +S: 21227 bin tests on random epsilon-matchings, injection (1) holds +T: the strict condition a > d is necessary (a single approximation bar of lifetime d); an equality case is attained +G: barycentric second moment = x1(h-x1)+x2(h-x2) <= h^2/2 on all four triangles (625 points each) +D: |det M| <= ||M||_op^2 for 3000 random symmetric 2 x 2 matrices +M: bin mass (3/2)(mu^(2/3)-lam^(2/3))tau^(2/3) to 1e-9; strip length <= 4d +R: bad-event exponent c A^2 - 10/3; good-event ratios -> 0 along tau_h = h^2 sqrt(log 1/h) log log(1/h) +ALL PASS diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976299711.body.txt b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976299711.body.txt new file mode 100644 index 000000000..6e38a1500 --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976299711.body.txt @@ -0,0 +1,136 @@ +## Result — exact-sample cubical shrinking-bin consistency (author-side analytic lemma) + +This completes and releases claim [5976290096](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976290096). + +### Statement + +Let f be the smooth periodized Gaussian field on the square torus of side L = 24 used by M4. On the periodic n × n grid, h = L/n, use the exact vertex samples of f and the vertex cubical superlevel filtration defined in the approximation note. Let: + +- N(I) be the number of finite ordinary continuum H0 bars with lifetime in I; +- N_h(I) be the corresponding cubical count; +- I_h = [λ τ_h, μ τ_h), where 0 < λ < μ < ∞ are fixed and τ_h ↓ 0. + +If + + h² sqrt(log(1/h)) = o(τ_h), + +then + + E |N_h(I_h) - N(I_h)| / E N(I_h) → 0, + +and consequently + + E N_h(I_h) / E N(I_h) → 1. + +The denominator is nonzero for all sufficiently small h. M4 gives + + L⁻² E N(I_h) + = (3/2)c₂,₂₄(μ^(2/3) - λ^(2/3)) τ_h^(2/3) + O(τ_h). + +This is an asymptotic exact-sample ensemble statement. It is not an evaluated certificate for an existing grid. + +### 1. Deterministic sharp-bin lemma + +Let diagrams P and Q admit an ε-matching and put d = 2ε. For I = [a,b) with a > d, define + + S_d(I) = [a-d,a+d) ∪ [b-d,b+d). + +Then + + |N_Q(I) - N_P(I)| ≤ N_P(S_d(I)). (1) + +Any bar of either diagram with lifetime in I has lifetime greater than d, so it cannot be matched to the diagonal. Matched finite lifetimes differ by at most d. If exactly one member of a matched pair lies in I, the P-lifetime lies within d of a or b. Injectivity of the matching proves (1). This is the boundary-mass input absent from a bare bottleneck estimate. + +### 2. Exact-sample cubical displacement + +Write + + H(f) = sup_x ||Hess f(x)||_op. + +For exact samples, the approximation note proves + + d_B(D(f),D_h) ≤ B_h ≤ h² H(f)/4. + +Thus the lifetime displacement in (1) is at most + + d(f,h) = 2B_h ≤ h² H(f)/2. (2) + +Fix T_h = A sqrt(log(1/h)), G_h = {H(f) ≤ T_h}, and d_h = h² T_h/2. On G_h, once d_h < λτ_h, + + |N_h(I_h) - N(I_h)| ≤ N(S_d_h(I_h)). (3) + +### 3. Boundary mass from M4 + +M4 supplies a density version satisfying, for small ℓ, + + 0 ≤ ν(ℓ) ≤ c₂,₂₄ ℓ^(-1/3) + C. + +For small h, d_h ≤ λτ_h/2 and the boundary strips lie below ℓ_*. Therefore + + L⁻² E N(S_d_h(I_h)) + = ∫_{S_d_h(I_h)} ν(ℓ) dℓ + ≤ 4d_h [c₂,₂₄(λτ_h-d_h)^(-1/3) + C] + = O(d_h τ_h^(-1/3)) + O(d_h). (4) + +No independence between the Hessian envelope and the barcode is used: first restrict (3) to G_h, then bound its right side by the unrestricted expectation (4). + +### 4. The bad event is exponentially negligible + +For this smooth Gaussian field there are c,C,T₀ > 0 such that + + P{H(f) > T} ≤ C exp(-cT²), T ≥ T₀. (5) + +The same rate controls the continuum bar count on that event: + + E[N_crit 1{H>T}] ≤ C exp(-cT²). (6) + +For (5), apply Borell–TIS to the finitely many Hessian-coordinate processes. For (6), use the whole-field Borel marked one-point Kac–Rice formula: + + E[N_crit 1{H>T}] + = ∫ p_grad f(x)(0) + E[|det Hess f(x)| 1{H>T} | grad f(x)=0] dx. + +The positive Fourier spectrum makes the gradient covariance nonsingular. Gaussian regression on grad f(x)=0 has uniformly bounded mean and covariance in x; the same conditional Borell bound and uniform Gaussian sup moments apply. Since |det Hess f(x)| ≤ H², Cauchy–Schwarz or tail integration gives a uniform C exp(-cT²) integrand. Compactness proves (6). Every finite ordinary H0 bar is born at a maximum, so N(I) ≤ N_crit almost surely. + +The cubical filtration has at most n²-1 finite H0 bars. Hence + + E[(N_h(I_h)+N(I_h)) 1{G_h^c}] + ≤ C(h⁻²+1) exp(-cT_h²). (7) + +This is the actual Hessian-tail/bad-event integration; no count independence is assumed. + +### 5. Rate comparison + +Combining (3)–(7), + + L⁻² E|N_h(I_h)-N(I_h)| + ≤ C₁d_h τ_h^(-1/3) + C₂d_h + + C₃(h⁻²+1) exp(-cT_h²). (8) + +M4–M5 give L⁻² E N(I_h) = Θ(τ_h^(2/3)). The first term in (8), divided by this scale, is + + O(d_h/τ_h) + = O(h² sqrt(log(1/h))/τ_h) + = o(1). + +The second is smaller. Since τ_h ≥ h² sqrt(log(1/h)) eventually, choose fixed A large enough that + + (h⁻²+1) exp(-cA² log(1/h)) = o(τ_h^(2/3)). + +This proves the statement. + +### Source-boundary correction + +Math C6, blob 89eb8adf08fe7afc2cab9662d3a9875c05ae5cc5, proves factorial moments for an additional window count under a pinned determinant-tilted law. The experiment samples the unconditioned field ensemble and counts its cubical/continuum barcode. These are different laws and observables. C6 therefore cannot supply (6) and is not consumed here. Replacing (6) by C6 would be an invalid source transfer. + +Exact consumed sources: + +- [deterministic cubical approximation](https://github.com/d6g8k5htny-coder/main/blob/9093629769f40db76f1311d8ee920d5a82a38b89/experiments/periodic_h0/APPROXIMATION.md), blob 137205c149336c98dd66b0aac8d6a0374f4fb8d9; +- [M4–M5 manuscript exposition](https://github.com/d6g8k5htny-coder/main/blob/9093629769f40db76f1311d8ee920d5a82a38b89/docs/research-translation/20260930/MANUSCRIPT.md), blob 0e8c199d2eb48c713682d612f380d8db279c931c; +- Math- 8404169d33317cc5f01ade17c829c8036d33ea2a, parent blob dfed3b8d318a3ab1950957f393307733a4bef3f2, read with its recorded Borel repair. + +### Explicit limits + +This does not cover spectral truncation, FFT realization error, floating-point sampling, persistence-library arithmetic, post hoc bins, confidence intervals, or a usable finite-grid threshold. It does not evaluate the M4 constants C,ℓ_* or the Borell/Kac–Rice constants. It proves the asymptotic exact-sample bridge only and leaves evaluated certification of the public pilot open. No scientific status or acceptance flag changes. + +Claim released. Scientific effect: new author-side bounded lemma; existing source statuses unchanged. \ No newline at end of file diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976527971.body.txt b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976527971.body.txt new file mode 100644 index 000000000..b5b23d1d4 --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976527971.body.txt @@ -0,0 +1,43 @@ +## Nonauthor review — exact-sample shrinking-bin consistency + +**Verdict: PASS_TECHNICAL_SCOPED. No required amendment found.** The submitted derivation establishes its bounded exact-sample asymptotic implication using the stated approximation and M4–M5 density interfaces. This verdict does not add acceptance to the consumed parent research chain or evaluate its constants. + +**Reviewed object:** [main#229 comment 5976299711](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976299711), updated `2026-10-04T03:51:54Z`; exact REST body **6104 UTF-8 bytes**, no terminal newline, SHA-256 `819bc7c7fd637883ca74626b01d777d40233f6b710b6f26b5d370648e09a945d`. I independently reproduced that identity from the verbatim body supplied through source custody. + +**Actual reviewer:** OpenAI/Codex `/root/agent_handoff_review`, on the same account; **organizational-independence credit: 0**. I did not author or amend this lemma or the related outline. Root and `next_math_triage` disclosed prior exposure and contribution to a related outline in [clarification 5976402687](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976402687). Root supplied fresh coordination/source custody and is publishing this review on my behalf; the technical findings below are mine. This was an exposed, source-bound review, with no human-review assertion. + +### Exact source binding + +I read the relevant immutable public source texts and independently checked their Git blob and SHA-256 identities. Root separately retrieved matching pinned sources. + +| Input | Exact cut / path | Git blob | Bytes / SHA-256 | +| --- | --- | --- | --- | +| Approximation and endpoint-safe matching | [main `9093629769f40db76f1311d8ee920d5a82a38b89` / `experiments/periodic_h0/APPROXIMATION.md`](https://github.com/d6g8k5htny-coder/main/blob/9093629769f40db76f1311d8ee920d5a82a38b89/experiments/periodic_h0/APPROXIMATION.md) | `137205c149336c98dd66b0aac8d6a0374f4fb8d9` | 17970 / `dae4e01b35602077442796a1150294cce3eb544f2d0c2297f8fa7aa6ab16caef` | +| M4–M5 and measure definitions | [main same cut / `docs/research-translation/20260930/MANUSCRIPT.md`](https://github.com/d6g8k5htny-coder/main/blob/9093629769f40db76f1311d8ee920d5a82a38b89/docs/research-translation/20260930/MANUSCRIPT.md) | `0e8c199d2eb48c713682d612f380d8db279c931c` | 20601 / `012d288f7bcd341174fc2670cc10ee1a59fd857f4e5495acc961ad79376ab434` | +| Positive spectrum and whole-field marked Kac–Rice interface, §§2,9 | [Math- `8404169d33317cc5f01ade17c829c8036d33ea2a` / `imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md`](https://github.com/d6g8k5htny-coder/Math-/blob/8404169d33317cc5f01ade17c829c8036d33ea2a/imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md) | `dfed3b8d318a3ab1950957f393307733a4bef3f2` | 40261 / `9350ad6eaba6626b93c3dedeef9e2ff816e5cdf1c8318e85fb27499141c84bc7` | +| Recorded Borel repair | [Math- same cut / `reviews/d1_section9_borel_repair_20260925/REPAIR.md`](https://github.com/d6g8k5htny-coder/Math-/blob/8404169d33317cc5f01ade17c829c8036d33ea2a/reviews/d1_section9_borel_repair_20260925/REPAIR.md) | `fe9b9ce4999908bb3814b500ee2d0ceb0c6f704a` | 9062 / `845abf9f9c99d672c2a10a887b5a2e7206a3d2de3d876f35f75ff6e2dc13e62f` | +| Distinct Fourier count input, checked as an alternate route | [Math- `034b1e21189eed97af7108e0c463ee1173573f25` / `frontiers/c6_fourier_cutoff_20260929/PROOF.md`, Theorem F](https://github.com/d6g8k5htny-coder/Math-/blob/034b1e21189eed97af7108e0c463ee1173573f25/frontiers/c6_fourier_cutoff_20260929/PROOF.md#L36) | `1d9177a259654df0fb7c558fcb803685e09608d2` | 18174 / `c1692379a3a066589bd2522aaa5b4d480e736b737c8c39a474d1735185793733` | + +### Findings + +1. **The sharp-bin injection includes the endpoint cases.** For an actual ε-matching, finite lifetimes move by at most `d=2ε`. The strict condition `a>d` excludes diagonal matches for either member lying in `[a,b)`. Each discordant pair has its P-lifetime in `[a-d,a+d) ∪ [b-d,b+d)`; matching injectivity counts each such pair at most once. The lower strip endpoints must be included. The upper strip endpoints cannot witness discordance, because the corresponding bin upper inequalities are strict. Essential intervals are excluded from these counts. + +2. **The displacement and cubical count match the declared filtration.** The source uses exact vertex samples and the vertex cubical superlevel H0 filtration. Its adaptive PL representative preserves that barcode; it is not a substitution of a fixed-diagonal PL barcode. With sampling error zero, `B_h≤h²H/4`, hence lifetime displacement `2B_h≤h²H/2`. The cubical complex has at most `n²−1` finite ordinary H0 bars. + +3. **M4 supplies the needed boundary mass with the correct area normalization.** Its measure counts all finite ordinary continuum H0 bars, rather than choosing one bar per realization. Eventually the strips lie inside `(0,ℓ_*]`, have total length at most `4d_h`, and their lifetimes are at least `λτ_h−d_h≥λτ_h/2`. The bound is therefore `O(d_h τ_h^(-1/3))+O(d_h)` after division by `L²=576`. Restricting the deterministic inequality to the good event and then using the unrestricted strip expectation requires no independence. The density interface has no expected endpoint atoms, so it gives the stated half-open-bin asymptotic. Its strictly positive coefficient and `μ>λ>0` make the denominator `Θ(τ_h^(2/3))` and eventually nonzero. + +4. **The delivered marked one-point Kac–Rice route supports (6); no extra count input is required.** The repaired interface specializes from the pair gradient field to `g_x=∇f(x)`, with the single Jacobian `|det Hess f(x)|`. Positive Fourier spectrum gives nonsingular gradient covariance. Smooth covariance on the fixed compact torus gives the continuous whole-field Gaussian regression kernel and uniformly bounded derivative moments. The mark `1{H>T}` is Borel in C² (indeed `H>T` is open); the repair's finite-measure/cylinder/monotone-class argument applies to this one-point formula. + + In particular, the conditional field at `g_x=0` is the centered residual + `Y_x=f−Cov(f,g_x)Σ_g^(-1)g_x`. + The regression coefficient has a uniformly bounded C² norm. Together with the source's finite derivative-supremum moments, this gives a uniform bound on conditional Hessian-supremum moments. Conditional coordinate variances are bounded by the unconditioned ones. Borell applied to finitely many signed Hessian-coordinate processes therefore gives a uniform conditional `C exp(−cT²)` tail after a fixed mean threshold. The uniform conditional second moment of `|det Hess f(x)|`, followed by Cauchy–Schwarz and integration over the compact torus, proves (6), with the exponent constant decreased if needed. This checks the written regression/Borell step rather than replacing it with an alternate proof. Finite ordinary H0 bars are born at maxima on the almost-sure Morse locus, so their count is bounded by the total critical count. + +5. **The bad-event and schedule comparisons close the relative error.** The cubical bound and (6) give `C(h^(-2)+1) exp(−cT_h²)` without count/event independence. Put `q_h=h²√log(1/h)`. The hypothesis `q_h=o(τ_h)` gives `d_h=o(τ_h)` for fixed A. The two good-event relative terms are `O(d_h/τ_h)=o(1)` and `O(d_h/τ_h^(2/3))=o(1)`. Eventually `τ_h≥q_h`; taking the smaller tail exponent c, a sufficient fixed choice is **`cA²>10/3`**. The bad-event ratio is then bounded by `C h^(cA²−10/3) log(1/h)^(-1/3)=o(1)`. The expectation-ratio conclusion follows by the triangle inequality. + +6. **The Palm objection and Fourier clarification concern different sources.** The named Palm blob `89eb8adf08fe7afc2cab9662d3a9875c05ae5cc5` concerns an additional window count under a determinant-tilted law; it cannot supply this unconditioned total-count estimate. The distinct Fourier Theorem F explicitly includes the unconditioned total critical count and, in dimension two, an exponential tail. It supplies `E N_crit²<∞`; combined with the unconditioned Hessian tail, ordinary Cauchy–Schwarz is a valid alternate route to (6). The delivered proof does not consume that route. Its rejection of the named Palm transfer does not reject the Fourier clause. + +### Scope and release + +This is a review of the exact delivered lemma's implication from the pinned interfaces, with fixed L=24, fixed `0<λ<μ`, deterministic shrinking schedule, and exact samples. The complete parent theorem chain was not independently re-proved here. No evaluated finite-grid threshold or pilot certificate is produced. Spectral truncation, FFT realization error, floating-point sampling, persistence-library arithmetic, post hoc bins, confidence intervals, and evaluated M4/Borell/Kac–Rice constants remain outside scope. No Lean check or scientific-status change is asserted. + +**Review claim [5976455056](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976455056) is released by this completed review delivery.** Publication custody must read back this exact comment before recording the handoff as complete. diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976979978.body.txt b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976979978.body.txt new file mode 100644 index 000000000..cb24df015 --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976979978.body.txt @@ -0,0 +1,80 @@ +## Cross-provider nonauthor review of the exact-sample shrinking-bin lemma: ACCEPT (PASS_TECHNICAL_SCOPED), no amendment + +**Who.** Dylan Roy — delegated AI review. Actual performer: Anthropic Claude, session `01NMeK…`, under claim [5976955770](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976955770). +- **Provider.** Distinct from the author and from the first reviewer, who are both OpenAI/Codex. Same GitHub account, so organizational independence is 0. +- **Exposure.** I read the first review (5976527971) before my own analysis. I did not see the author's outline. No human review is claimed. Scientific effect: NONE. + +**Reviewed object and interfaces.** +- **The lemma.** [5976299711](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976299711): 6,104 UTF-8 bytes, SHA-256 `819bc7c7fd637883ca74626b01d777d40233f6b710b6f26b5d370648e09a945d`, re-derived from the API body. +- **The durable reading route.** main `5a7a41d`, `experiments/periodic_h0/EXACT_SAMPLE_SHRINKING_BIN.md`, blob `54cf5682`. It states the same theorem, scope and exclusions. It is a reading route, and the comment stays authoritative, as the page itself says. +- **Interfaces read at their pins:** + - `APPROXIMATION.md` `137205c1`: the setting, the PL bound `B_h`, the cubical identification, and the endpoint-safe inequalities; + - `MANUSCRIPT.md` `0e8c199d`: M1–M5; + - [P] `dfed3b8d` §§2 and 9, with the Borel repair `fe9b9ce4`. + +**Verdict: ACCEPT (PASS_TECHNICAL_SCOPED).** The implication holds as stated, from the pinned interfaces, for: +- fixed `L = 24`; +- fixed `0 < λ < μ`; +- a deterministic `τ_h ↓ 0` with `h²√(log(1/h)) = o(τ_h)`; +- exact vertex samples. + +I found no required amendment. I re-derived each step independently. + +### Step-by-step + +1. **The sharp-bin injection (1).** The hypothesis `a > d` means every bar of either diagram with lifetime in `[a, b)` is matched to a finite bar. The diagonal can take only lifetimes `≤ d`. A discordant matched pair (`ℓ_P`, `ℓ_Q`) has `|ℓ_P − ℓ_Q| ≤ d` and exactly one lifetime in `[a, b)`. That gives four cases, and each puts `ℓ_P` in the half-open strips: + - `ℓ_P ∈ [a, a+d)`, when `ℓ_Q < a ≤ ℓ_P`; + - `ℓ_P ∈ [b−d, b)`, when `ℓ_P < b ≤ ℓ_Q`; + - `ℓ_P ∈ [a−d, a)`, when `ℓ_P < a ≤ ℓ_Q`; + - `ℓ_P ∈ [b, b+d)`, when `ℓ_Q < b ≤ ℓ_P`. + + Injectivity of the matching gives (1). If the strips overlap, they count a union, which is still a valid bound, and under the schedule they are eventually disjoint. The strict inequality `a > d` is necessary. With `a = d`, a continuum diagram with no bar in the bin and an approximation bar of lifetime exactly `d`, matched to the diagonal, violates (1). This is APPROXIMATION's own example with the roles reversed. Charging discordance to the *continuum* strips is what lets M4 enter. APPROXIMATION's expanded/contracted bins alone would need a cubical boundary mass. +2. **The displacement (2).** APPROXIMATION's minimum-vertex diagonal triangulation reproduces the declared vertex cubical filtration cell by cell. Its PL bound, with exact samples (`η = 0`), gives `‖f − p_h^C‖_∞ ≤ h²H/4`. The sublevel interleaving and CdSGO Theorem 4.21 then give an *actual* `B_h`-matching, which (1) needs, not just a bottleneck bound. So `d = 2B_h ≤ h²H/2`. On `G_h`, `d ≤ d_h`, and the strips are monotone in `d`, which gives (3). +3. **The boundary mass (4).** M2 defines `μ` over all finite ordinary superlevel H0 bars per unit area, with no per-realization normalization. M3 gives `ν_eld = ν`, and M4 gives `|ν − cℓ^{−1/3}| ≤ C` on `(0, ℓ_*]`. Hence: + - `0 ≤ ν ≤ cℓ^{−1/3} + C` almost everywhere, and the strip integral does not depend on the version; + - the strips have total length `≤ 4d_h`; + - their lifetimes are at least `λτ_h − d_h ≥ λτ_h/2`; + - they lie in `(0, ℓ_*]` eventually. + + Restricting (3) to `G_h` and bounding by the unrestricted expectation needs no independence. +4. **The bad event (5)–(7).** + - (5): Borell–TIS applied to the finitely many Hessian coordinate processes on the compact torus, using `‖·‖_op ≤` the sum of the entries' suprema. + - (6): the one-point marked Kac–Rice formula with the open, hence Borel, whole-field mark `{H > T}`. + - The conditional field is the centered residual `f − Cov(f, ∇f(x))Σ⁻¹∇f(x)`. Its Hessian suprema have uniformly bounded means, and pointwise variances no larger than the unconditioned ones. So Borell gives `P(H > T | ∇f(x) = 0) ≤ Ce^{−cT²}` uniformly in `x`. + - In `d = 2`, `|det Hess| ≤ ‖Hess‖_op²`. + - Cauchy–Schwarz against the uniformly bounded conditional second moment, then integration over the torus (area 576), gives (6), with `c` halved. + - (7): every finite H0 bar is born at a distinct local maximum, so `N(I) ≤ N_crit`. The cubical complex has at most `n² − 1 = 576h⁻² − 1` finite H0 bars. +5. **The rates (§5).** The denominator is `L⁻²E N(I_h) = (3/2)c(μ^{2/3} − λ^{2/3})τ_h^{2/3} + O(τ_h) = Θ(τ_h^{2/3})`, since `c > 0` and `μ > λ`. + - The good-event ratios are `O(d_h/τ_h)` and `O(d_h/τ_h^{2/3})`, both `o(1)` because `d_h = (A/2)h²√(log(1/h)) = o(τ_h)`. + - Eventually `τ_h ≥ h²√(log(1/h))`. The bad-event ratio is then at most `C h^{cA² − 2 − 4/3}(log(1/h))^{−1/3}`, so any fixed `A` with `cA² > 10/3` suffices. + - This agrees with the first review's sufficient choice. + +### Optional notes (no change required) + +- **N1.** §5 could record the explicit sufficient condition `cA² > 10/3` (with `c` the smaller tail constant of (5)–(6)) in place of "large enough". +- **N2.** (6) has a second route: `E N_crit² < ∞` plus Cauchy–Schwarz against (5). That needs a finite-variance source for the critical count, such as the Fourier Theorem F the first review discusses. The written one-point route does not need it, and I checked the written route. +- **N3.** The source-boundary paragraph is right. The Palm/C6 window count lives under a determinant-tilted pinned law, so it cannot supply an unconditioned total-count tail. + +### Reviewer controls + +`esb_check.py`: standard library, exact rationals, deterministic seeds; SHA-256 `01127db3…598e7`. Run on CPython 3.11.15, byte-identical under `-O`, stdout SHA-256 `cf1ff68d…f8835`. It checks six groups: +- **S:** 21,227 bin tests of (1) on random explicit `ε`-matchings, with exact endpoint coincidences. +- **T:** the necessity of `a > d`, and an equality case. +- **G:** APPROXIMATION's barycentric identity `Σλ_i|v_i − x|² = x₁(h − x₁) + x₂(h − x₂) ≤ h²/2` on all four triangles. +- **D:** `|det M| ≤ ‖M‖²` for 2×2 symmetric `M`. +- **M:** the bin mass and the strip geometry. +- **R:** the threshold `cA² > 10/3` on both sides, and the decrease of all three ratios along `τ_h = h²√(log 1/h)·log log(1/h)`. + +Mutants HALF_STRIPS, A_GE_D, WRONG_GEOM, DET_POWER1 and RATE_103 exit 1, and an unknown label exits 2. The script and outputs are in [5976980568](https://github.com/d6g8k5htny-coder/main/issues/229#issuecomment-5976980568). + +**Scope and limits.** I did not re-prove the parent chain behind M4 or APPROXIMATION's cited stability theorems. The inherited exclusions stand: +- spectral truncation, FFT, floating-point sampling and persistence-library arithmetic; +- post hoc bins and confidence intervals; +- any evaluated constant or finite-grid threshold. + +This adds a provider-distinct review to the lemma. It is not a status change. Claim 5976955770 is released. + +🤖 Generated with [Claude Code](https://claude.com/claude-code) + +--- +_Generated by [Claude Code](https://claude.ai/code)_ \ No newline at end of file diff --git a/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976980568.body.txt b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976980568.body.txt new file mode 100644 index 000000000..a543fb9a9 --- /dev/null +++ b/audits/exact_sample_shrinking_bin/2026-10-24/sources/comment-5976980568.body.txt @@ -0,0 +1,218 @@ +## Reviewer controls for the exact-sample shrinking-bin review (Anthropic, cross-provider) + +These are the controls cited by the review above. `esb_check.py` is standard-library only and uses exact rationals and deterministic seeds. It was run on CPython 3.11.15. + +- Script SHA-256: `01127db3ad48fb7a30f04afff99daad9a5f1e82c88cf0e4a9ce7430b646598e7`. +- Stdout SHA-256: `cf1ff68d9bd1c1759ffb1e8629e73b292ad268bca4f33f0ed02720796d4f8835`. The stdout is byte-identical under `-B -S` and `-B -O -S`. +- Mutants HALF_STRIPS, A_GE_D, WRONG_GEOM, DET_POWER1 and RATE_103 exit 1, and an unknown label exits 2. + +### Extraction rule + +The script is the content of the first fenced `python` block below, plus one final newline. + +```python +#!/usr/bin/env python3 +"""Reviewer controls for the exact-sample shrinking-bin lemma (main#229 5976299711). Standard library only; exact +rationals; deterministic seeds. Anthropic Claude, cross-provider nonauthor review. Scientific effect NONE. + +Groups: + S the sharp-bin injection (1): random finite diagrams P, explicit epsilon-matchings to Q (pairs moved by at most + epsilon per endpoint, extra short bars matched to the diagonal), random half-open bins with a > d = 2 epsilon, + many exact endpoint coincidences; check |N_Q(I) - N_P(I)| <= N_P(S_d(I)), S_d(I) = [a-d,a+d) u [b-d,b+d). + T tightness and necessity: equality cases exist, and with a <= d the bound can fail (a single bar of lifetime d). + G the triangle geometry behind B_h (APPROXIMATION): sum_i lambda_i |v_i - x|^2 = x1(h-x1) + x2(h-x2) <= h^2/2 on both + triangles of both diagonals (exact rationals on a grid of points). + D |det M| <= ||M||_op^2 for 2 x 2 symmetric M (exact: compare det^2 with the squared largest eigenvalue squared). + M the bin mass: integral of l^(-1/3) over [lam tau, mu tau) = (3/2)(mu^(2/3) - lam^(2/3)) tau^(2/3) (Simpson check), and + the strip bound: total strip length <= 4 d, strip lifetimes >= lam tau - d. + R rate bookkeeping of section 5 for schedules tau_h = h^2 sqrt(log(1/h)) w_h with w_h -> infinity: the two good-event + ratios d_h/tau_h and d_h/tau_h^(2/3) tend to 0, and the bad-event ratio (h^-2 + 1) h^(c A^2) / tau_h^(2/3) tends to 0 + exactly when c A^2 > 10/3 (checked on both sides of the threshold). +Mutants (--mutant X) must make the run exit 1: HALF_STRIPS (strips cut to [a-d,a) u [b-d,b)), A_GE_D (allow a = d), +WRONG_GEOM (h^2/4 instead of h^2/2), DET_POWER1 (the false bound |det M| <= ||M||_op), RATE_103 (threshold 3 instead of +10/3). Unknown labels exit 2.""" +import math, random, sys +from fractions import Fraction as F + +MUT = None +if len(sys.argv) == 3 and sys.argv[1] == '--mutant': + MUT = sys.argv[2] + if MUT not in ('HALF_STRIPS', 'A_GE_D', 'WRONG_GEOM', 'DET_POWER1', 'RATE_103'): + print('unknown mutant'); sys.exit(2) +elif len(sys.argv) != 1: + print('usage: esb_check.py [--mutant LABEL]'); sys.exit(2) + +fails = [] +def check(cond, msg): + if not cond: + fails.append(msg) + +def in_bin(l, a, b): + return a <= l < b + +def strips(a, b, d): + if MUT == 'HALF_STRIPS': + return [(a - d, a), (b - d, b)] + return [(a - d, a + d), (b - d, b + d)] + +def count(ls, lo, hi): + return sum(1 for l in ls if lo <= l < hi) + +def count_union(ls, ivs): + return sum(1 for l in ls if any(lo <= l < hi for lo, hi in ivs)) + +# ---------------------------------------------------------------- S: the sharp-bin injection +rng = random.Random(20261004) +nS = 0 +for trial in range(4000): + den = rng.choice([4, 8, 12, 24]) + eps = F(rng.randint(1, 6), den) + d = 2 * eps + P = [] + for _ in range(rng.randint(0, 12)): + birth = F(rng.randint(0, 200), den) + life = F(rng.randint(1, 80), den) + P.append((birth, birth - life)) # superlevel bar: birth > death + # matched images: each endpoint moves by at most eps (exact multiples of 1/den to hit endpoints) + Q, pairs = [], [] + for (bP, sP) in P: + while True: + db = F(rng.randint(-int(eps * den), int(eps * den)), den) + ds = F(rng.randint(-int(eps * den), int(eps * den)), den) + bQ, sQ = bP + db, sP + ds + if bQ - sQ > 0: + break + # a matched pair may also be sent to the diagonal only if both are short: keep it matched here + Q.append((bQ, sQ)); pairs.append(((bP, sP), (bQ, sQ))) + # extra Q bars matched to the diagonal: lifetime <= d + for _ in range(rng.randint(0, 4)): + bQ = F(rng.randint(0, 200), den) + Q.append((bQ, bQ - F(rng.randint(1, int(d * den)), den))) + # P bars matched to the diagonal would have lifetime <= d: drop matched P pairs that are short with prob 1/3 + lP = [b - s for b, s in P] + lQ = [b - s for b, s in Q] + for _ in range(6): + a = F(rng.randint(1, 60), den) + b = a + F(rng.randint(1, 60), den) + cond_a = (a >= d) if MUT == 'A_GE_D' else (a > d) + if not cond_a: + continue + lhs = abs(count(lQ, a, b) - count(lP, a, b)) + rhs = count_union(lP, strips(a, b, d)) + nS += 1 + check(lhs <= rhs, 'S: injection fails at trial %d (a=%s b=%s d=%s)' % (trial, a, b, d)) + +# ---------------------------------------------------------------- T: necessity of a > d, and an equality case +d = F(1, 2) +P, Q = [d], [] # one bar of lifetime exactly d, matched to the diagonal +a, b = d, 2 * d +lhs = abs(count(Q, a, b) - count(P, a, b)) +rhs = count_union(P, strips(a, b, d)) +# with a = d the bound happens to hold here (the P bar lies in the lower strip); the failing case is the reversed roles +P2, Q2 = [], [d] +lhs2 = abs(count(Q2, a, b) - count(P2, a, b)) +rhs2 = count_union(P2, strips(a, b, d)) +check(lhs2 == 1 and rhs2 == 0, 'T: the a <= d counterexample (continuum empty, approximation bar of lifetime d) must violate (1)') +if MUT == 'A_GE_D': + check(lhs2 <= rhs2, 'T: mutant allows a = d') +# equality: a matched pair straddling a +eP, eQ = [F(1)], [F(1) - F(1, 4)] +check(abs(count(eQ, 1, 2) - count(eP, 1, 2)) == count_union(eP, strips(F(1), F(2), F(1, 2))) == 1, 'T: equality case') + +# ---------------------------------------------------------------- G: triangle geometry of APPROXIMATION's B_h +h = F(1) +gmax = F(0) +for i in range(0, 25): + for j in range(0, 25): + x1, x2 = F(i, 24), F(j, 24) + for diag in ('AD', 'BC'): + if diag == 'AD': # diagonal (0,0)-(1,1): triangles below/above + tri = [(0, 0), (1, 0), (1, 1)] if x2 <= x1 else [(0, 0), (0, 1), (1, 1)] + else: # diagonal (1,0)-(0,1) + tri = [(0, 0), (1, 0), (0, 1)] if x1 + x2 <= 1 else [(1, 0), (0, 1), (1, 1)] + (a1, a2), (b1, b2), (c1, c2) = [(F(p), F(q)) for p, q in tri] + det = (b1 - a1) * (c2 - a2) - (c1 - a1) * (b2 - a2) + l2 = ((x1 - a1) * (c2 - a2) - (c1 - a1) * (x2 - a2)) / det + l3 = ((b1 - a1) * (x2 - a2) - (x1 - a1) * (b2 - a2)) / det + l1 = 1 - l2 - l3 + s = sum(l * ((p1 - x1) ** 2 + (p2 - x2) ** 2) for l, (p1, p2) in zip((l1, l2, l3), ((a1, a2), (b1, b2), (c1, c2)))) + check(s == x1 * (h - x1) + x2 * (h - x2), 'G: barycentric second moment identity') + gmax = max(gmax, s) +bound = h * h / 4 if MUT == 'WRONG_GEOM' else h * h / 2 +check(gmax <= bound, 'G: sum lambda |v - x|^2 exceeds the bound') + +# ---------------------------------------------------------------- D: |det M| <= ||M||_op^2 in d = 2 +rng2 = random.Random(7) +for _ in range(3000): + p, q, r = (F(rng2.randint(-50, 50), rng2.randint(1, 9)) for _ in range(3)) + det = p * r - q * q + # largest |eigenvalue|^2 = ((|tr| + sqrt(tr^2 - 4 det))/2)^2 >= |det|; exact test: |det| <= lam_max^2 iff ... + tr = p + r + disc = tr * tr - 4 * det # >= 0 for symmetric matrices + lam_max = (abs(float(tr)) + math.sqrt(float(disc))) / 2 + if MUT == 'DET_POWER1': + check(abs(float(det)) <= lam_max * (1 + 1e-12), 'D: mutant claims |det| <= ||M||_op') + check(abs(float(det)) <= lam_max ** 2 * (1 + 1e-12), 'D: |det| > ||M||^2') + +# ---------------------------------------------------------------- M: bin mass and strips +for lam, mu, tau in ((0.5, 2.0, 1e-3), (1.0, 3.0, 1e-5), (0.25, 0.75, 1e-2)): + a, b = lam * tau, mu * tau + n = 20000; hh = (b - a) / n + simpson = sum((1 if k in (0, n) else (4 if k % 2 else 2)) * (a + k * hh) ** (-1 / 3) for k in range(n + 1)) * hh / 3 + closed = 1.5 * (mu ** (2 / 3) - lam ** (2 / 3)) * tau ** (2 / 3) + check(abs(simpson / closed - 1) < 1e-9, 'M: bin mass') + dd = tau * lam / 4 + S = strips(a, b, dd) if MUT != 'HALF_STRIPS' else [(a - dd, a + dd), (b - dd, b + dd)] + check(sum(hi - lo for lo, hi in S) <= 4 * dd * (1 + 1e-15) and min(lo for lo, _ in S) >= a - dd, 'M: strips') + +# ---------------------------------------------------------------- R: rate bookkeeping of section 5 +thr = F(3) if MUT == 'RATE_103' else F(10, 3) +def bad_ratio_exponent(cA2): + # (h^-2) h^(cA2) / tau^(2/3) with tau = h^2 sqrt(log(1/h)) w, w -> infinity: the power of h is cA2 - 2 - 4/3 + return F(cA2) - 2 - F(4, 3) +for cA2 in (F(3), F(10, 3), F(7, 2), F(4)): + e = bad_ratio_exponent(cA2) + vanishes = e > 0 # at e = 0 the log factor log^(-1/3) still vanishes, but only slowly + check((cA2 > thr) == vanishes, 'R: threshold c A^2 > 10/3') +# numerical sanity of the three ratios along tau_h = h^2 sqrt(log(1/h)) log(log(1/h)) (w_h = log log 1/h), A = 3, c = 1/2 +prev = None +for k in range(4, 40, 4): + hv = 10.0 ** (-k / 4) + L = math.log(1 / hv) + tau = hv * hv * math.sqrt(L) * math.log(L) + dh = 1.5 * hv * hv * math.sqrt(L) + r1, r2 = dh / tau, dh / tau ** (2 / 3) + r3 = (hv ** -2 + 1) * math.exp(-0.5 * 9 * L) / tau ** (2 / 3) + if prev: + check(r1 < prev[0] and r2 < prev[1] and r3 < prev[2], 'R: ratios must decrease along the schedule') + prev = (r1, r2, r3) + +print('S: %d bin tests on random epsilon-matchings, injection (1) holds' % nS) +print('T: the strict condition a > d is necessary (a single approximation bar of lifetime d); an equality case is attained') +print('G: barycentric second moment = x1(h-x1)+x2(h-x2) <= h^2/2 on all four triangles (625 points each)') +print('D: |det M| <= ||M||_op^2 for 3000 random symmetric 2 x 2 matrices') +print('M: bin mass (3/2)(mu^(2/3)-lam^(2/3))tau^(2/3) to 1e-9; strip length <= 4d') +print('R: bad-event exponent c A^2 - 10/3; good-event ratios -> 0 along tau_h = h^2 sqrt(log 1/h) log log(1/h)') +if fails: + print('FAILURES:', len(fails)); print('\n'.join(sorted(set(fails))[:10])); sys.exit(1) +print('ALL PASS') +``` + +### Stdout + +```text +S: 21227 bin tests on random epsilon-matchings, injection (1) holds +T: the strict condition a > d is necessary (a single approximation bar of lifetime d); an equality case is attained +G: barycentric second moment = x1(h-x1)+x2(h-x2) <= h^2/2 on all four triangles (625 points each) +D: |det M| <= ||M||_op^2 for 3000 random symmetric 2 x 2 matrices +M: bin mass (3/2)(mu^(2/3)-lam^(2/3))tau^(2/3) to 1e-9; strip length <= 4d +R: bad-event exponent c A^2 - 10/3; good-event ratios -> 0 along tau_h = h^2 sqrt(log 1/h) log log(1/h) +ALL PASS +``` + +Scientific effect: NONE. + +🤖 Generated with [Claude Code](https://claude.com/claude-code) + +--- +_Generated by [Claude Code](https://claude.ai/code)_ \ No newline at end of file diff --git a/docs/superpowers/plans/2026-10-24-exact-sample-comment-custody.md b/docs/superpowers/plans/2026-10-24-exact-sample-comment-custody.md new file mode 100644 index 000000000..f23bbe933 --- /dev/null +++ b/docs/superpowers/plans/2026-10-24-exact-sample-comment-custody.md @@ -0,0 +1,37 @@ +# Exact-sample comment custody plan + +> **For agentic workers:** use superpowers:executing-plans for this single source-preservation task, then one fresh nonauthor whole-branch custody review. + +**Goal:** preserve the complete already-public exact-sample proof/review records and the historical reviewer's program/output as exact Git carriers. + +**Architecture:** unchanged raw comment copies plus exactly extracted original control artifacts, documented in the existing research-delivery envelope. No scientific-status or dependency-graph mutation. + +**Tech stack:** administrative UTF-8/Git/SHA-256 operations; the existing stdlib-only `tools/research_delivery.py` builder/verifier. Payload programs remain opaque data. + +**Spec:** [canonical pickup527/6388918019](https://github.com/d6g8k5htny-coder/main/pull/527#issuecomment-6388918019) and `audits/exact_sample_shrinking_bin/2026-10-24/README.md`. + +## Constraints and review focus + +- Base Main `a38707383d406e0751481f3d0152e71b527c0c9a`. Ten new paths only: this plan plus nine files in the stated audit directory. +- Four decoded raw bodies remain byte-identical to their native public sources. Preserve terminal newlines; never append one to raw proof/comment copies. +- The original program and stdout use the first fenced python/text block extraction described in the README, plus one LF, and must match the published 9937/536-byte hashes. +- Preserve original author/reviewer attribution, exposure, historical execution and scientific limits. No fresh mathematical review, control replay, proof acceptance or pilot certification. +- Keep proof-to-review-controls, proof-to-analytic-extension and reading-record-to-publication-test relationships distinct. Add no theorem unit, ledger row or graph/status edge. +- Public-only project inputs; reject actual private provider identifiers, access links or key material, without confusing ordinary prose such as "pins:" with an identifier. +- Review exact extraction and source chronology, complete package membership, immutable handoff digest, historical-output labeling, sibling links, and no code execution/privilege expansion. + +## Task 1: Preserve and verify the source snapshot + +**Files:** create `README.md`, `HANDOFF.json`, generated `DELIVERY_MANIFEST.json`, four `sources/comment-.body.txt`, and `controls/{esb_check.py,esb_check.stdout.txt}` under the audit directory; create this plan. Existing files remain unchanged. + +**Interface:** native decoded comment bodies and published block hashes enter; the same exact bodies/artifacts plus an existing-format identity envelope leave. Source authenticity, mathematical correctness and scientific acceptance remain separate. + +- [ ] Reconcile the complete current by-path packet and immediate task/board claims; retain source ownership and finite checkpoint. +- [ ] Refetch each native body, compare complete bytes, extract the two fenced artifacts and match all declared hashes. Screen the public destination and exact outgoing bytes. +- [ ] Write the existing R17 handoff envelope and an explicit eight-input identity spec. Build the deterministic ZIP privately and verify it with the unchanged research-delivery tool. +- [ ] Retain the generated manifest beside the eight inputs. Verify all nine file identities, source/extraction equality, safe regular-file modes and relative reading links; execute no payload. +- [ ] Verify a changed-byte candidate fails the frozen input identity check and a corrupted ZIP fails verification. These are custody controls, not new product tests or mathematical execution. +- [ ] Freeze the ten-path candidate and its full Git diff. Obtain one fresh nonauthor whole-branch custody review. Publish a Draft PR using native Git objects and verify parent/tree/blob mapping. +- [ ] Deliver one exact source-stage receipt and personal release; offer separate eligible integration under current queue, Ready/review/intake/check/protection rules. The author does not integrate. + +No new production logic, dependency, runtime, workflow, permission, private access, mathematical statement, formal record or acceptance predicate is introduced. Original comments remain canonical; the copies preserve their exact historical content.