From 2fa23bace0ec25afa48e76c5a374a25c68d5345a Mon Sep 17 00:00:00 2001 From: Claude Date: Tue, 29 Sep 2026 18:10:00 +0000 Subject: [PATCH 1/3] Add C7-total reconnaissance: gap-mark scaling of selection failure, l^(-1/4) conjecture MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit A reconnaissance memo (proves nothing) for the open C7-total question: the order of the unrestricted rejected density rho_rej = nu_cand^all - nu_eld^all, currently known only as o(l^(-1/3)) with Theta(1) far part. Tracking the gap mark k through the parent's own §§5-7 shows the depth- failure layer has width ~ r/k (D = 4K^2/(3k) in (7.1)), so the layer's weighted mass is ~ (r/k)^3 (7.3) — provided the axial factors are kept at their true size 6k (P (5.2)) instead of the canceled-pivot bound M_3/2 (6.1), since the normalizer Z_r ~ k^2 r^2 (5.4) carries the same k^2. Heuristic (H): 1 - p_r <= C min{1, (r/k)^3} (b-tailed). Inserting (H) into the parent's (13.5) with r^3 = l/k gives threshold k_* = l^(1/4) and rho_rej^near = O(l^(-1/4)); the rejected measure then has finite inverse moments for q > -3/4, splitting the open strip (-1, -2/3] at -3/4. Conjecture C7-tau: rho_rej(l) ~ l^(-1/4) in every d; the unrestricted difference is NOT O(1). The lower half is genuinely open: the planar/#149 direct events yield only c k^6 r^3 at small k (they pin O(k) geometry); the conjectured dominant layer lambda_1 ~ r M_3^2/k with O(1) third derivatives has ledger (r/k)^3 and needs its own failure mechanism. Three ordered proof obligations are listed. exponent_check.py verifies the piecewise power-law integrals and both ledgers exactly (stdlib, -O identical, two mutants exit 1). Consistent with every accepted result. Scientific effect: NONE; no register change; not a theorem; no self-merge. Co-Authored-By: Claude Fable 5.1 Claude-Session: https://claude.ai/code/session_01NMeKEismAyeqgdB4sy2NJU --- .../workflows/c7-total-k-scaling-recon.yml | 59 ++++++++++++ .../RECONNAISSANCE.md | 96 +++++++++++++++++++ .../RESULTS.json | 23 +++++ .../SOURCE_FILES.json | 30 ++++++ .../exponent_check.py | 80 ++++++++++++++++ 5 files changed, 288 insertions(+) create mode 100644 .github/workflows/c7-total-k-scaling-recon.yml create mode 100644 reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md create mode 100644 reviews/c7_total_k_scaling_recon_20260929/RESULTS.json create mode 100644 reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json create mode 100644 reviews/c7_total_k_scaling_recon_20260929/exponent_check.py diff --git a/.github/workflows/c7-total-k-scaling-recon.yml b/.github/workflows/c7-total-k-scaling-recon.yml new file mode 100644 index 000000000..b65c1ee0c --- /dev/null +++ b/.github/workflows/c7-total-k-scaling-recon.yml @@ -0,0 +1,59 @@ +name: C7-total k-scaling reconnaissance exponent checks + +'on': + pull_request: + paths: + - 'reviews/c7_total_k_scaling_recon_20260929/**' + - '.github/workflows/c7-total-k-scaling-recon.yml' + workflow_dispatch: + +permissions: + contents: read + +jobs: + exact-and-probe: + runs-on: ubuntu-latest + timeout-minutes: 10 + steps: + - uses: actions/checkout@11d5960a326750d5838078e36cf38b85af677262 + with: + persist-credentials: false + - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 + with: + python-version: '3.11.16' + - name: Verify sources, replay both modes, reject mutants + env: + PYTHONDONTWRITEBYTECODE: '1' + run: | + set -euo pipefail + python -B -S - <<'PY' + import hashlib, json, pathlib, subprocess, sys + root = pathlib.Path('reviews/c7_total_k_scaling_recon_20260929').resolve() + manifest = json.loads((root / 'SOURCE_FILES.json').read_text())['files'] + present = sorted(p.name for p in root.iterdir()) + expected_names = sorted([e['path'] for e in manifest] + ['SOURCE_FILES.json']) + if present != expected_names: + raise ValueError('packet tree differs from manifest: ' + repr(present)) + for entry in manifest: + path = root / entry['path'] + if pathlib.PurePosixPath(entry['path']).name != entry['path'] or path.is_symlink(): + raise ValueError('flat regular source required: ' + entry['path']) + data = path.read_bytes() + if len(data) != entry['bytes'] or hashlib.sha256(data).hexdigest() != entry['sha256']: + raise ValueError('source identity mismatch: ' + entry['path']) + script = str(root / 'exponent_check.py') + expected = (root / 'RESULTS.json').read_bytes() + mutants = ['M1', 'M2'] + for flags in (['-B', '-S'], ['-B', '-O', '-S']): + out = subprocess.run([sys.executable, *flags, script], + check=True, capture_output=True, timeout=300).stdout + if out != expected: + raise RuntimeError('output differs from RESULTS.json under ' + ' '.join(flags)) + for m in mutants: + rc = subprocess.run([sys.executable, *flags, script, '--mutant', m], + capture_output=True, timeout=300).returncode + if rc != 1: + raise RuntimeError('mutant ' + m + ' not rejected') + print(json.dumps({'scientific_effect': 'NONE', 'mathematical_acceptance': False})) + PY + git diff --exit-code diff --git a/reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md b/reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md new file mode 100644 index 000000000..1b87bae9a --- /dev/null +++ b/reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md @@ -0,0 +1,96 @@ +# Reconnaissance memo — C7-total: the gap-mark scaling of selection failure and the `ℓ^{−1/4}` conjecture — 2026-09-29 + +**Object:** CL-C7-TOTAL-K-SCALING-RECON-20260929-v1. **Author:** Anthropic Claude (session `session_01NMeKEismAyeqgdB4sy2NJU`). +**Scientific effect:** NONE. **This memo proves nothing.** It records a +scaling analysis of the parent's own proof with the gap mark `k` tracked, states the conjecture it points to, and lists +the exact proof obligations, so that a lane attacking C7-total attacks the right object. + +## 1. The open question, as the C7 record leaves it + +`reviews/c7_nonvanishing_openai_20260929/REVIEW.md` §5: far rejected pairs have density `Θ(1)` (Theorem U lower, parent +(14.1) upper); the **total** unrestricted rejected density `ρ_rej(ℓ) := ν_cand^all(ℓ) − ν_eld^all(ℓ)` is known only to be +`o(ℓ^{−1/3})`; the inverse-moment strip `q ∈ (−1, −2/3]` for the rejected measure is unresolved; "no unrestricted `O(1)` +conclusion follows." + +## 2. Where the near rejected density comes from + +By the parent's (13.5) applied to the candidate/elder difference (`1 − p_r` in place of `p_r`), with `r = (ℓ/k)^{1/3}`, + + ρ_rej^near(ℓ) = ℓ^{−1/3} ∫_{ℝ×(0,∞)×S^{d−1}} 1{k ≥ ℓ/r_0³} [1 − p_r(b,k,u)] A_r(b,k,u) / (3k^{2/3}) db dk dσ(u). + +Everything is in how `1 − p_r` depends on `k` as `k ↓ 0`. The parent's Theorem A is stated for compact `K ⊂ (0,∞)` and +deliberately never tracks `k` (P §13: "Do NOT require … a globally uniform constant in (1.1)"). + +## 3. Tracking `k` through P §§5–7: the selection bound is `min{1, (r/k)³}`, heuristically + +Read P §7 with `k` explicit: + +- The cap region is `G_r = {λ_min(−A_M) > (4/(3k)) r M_3², rM_4 ≤ 3k/10}`. So depth failure is the layer + `0 < λ_1 ≤ D r U²` with `D = 4K²/(3k)` (P (7.1)): **the layer width is `∝ r/k`**. +- P (7.3) integrates the soft factor over that layer: `∫_0^{DrU²} λ_1(λ_1 + ErU) dλ_1 ≍ (Dr)³ ≍ (r/k)³` (times moments of + `U`). Combined with the `r²` from the other factors, the numerator `E_Q[W_r 1_{fail}]` is `≍ r² (r/k)³` — **if** the + axial factors are kept at their true size. +- P (6.1)–(6.2) bound `|α_M|, |α_S|` by `h/2 = M_3/2` (canceled-pivot). But P (5.2) says `α_M ≈ −6k`, `α_S ≈ 6k` when + `rM_4 ≪ k`. Bounding `6k` by `M_3/2` loses a factor `k²` in the numerator that the denominator does not lose: + P (5.4) gives `Z_r/r² → z_0 = (6k)² E[det(A_0)² 1{A_0<0}]`, i.e. **`Z_r ≍ k² r²`**. With the true axial size the `k²` + cancels between numerator and normalizer; with the canceled-pivot bound it does not. +- The fourth-derivative exception P (7.7) is `[10r/(3k)]⁴ E[(W/r²) M_4⁴] ≍ (r/k)⁴`, subdominant to `(r/k)³` for `r < k`. + +Heuristic conclusion (H): `1 − p_r(b,k,u) ≲ C min{1, (r/k)³} · (1 + |b| + k)^N` uniformly, for `r ≤ r_0`. +For `k` in a compact window this is the parent's `C r³`; the new content is the `k^{−3}` growth as `k ↓ 0` and the +saturation at `r ≍ k`. Physically: the whole cubic pair landscape `G_k` is `O(k)` in scaled units (every term of +`G_k` carries `k`), while third-order fluctuations are `O(1)`; when `k ≪ 1` the pairing survives only if the +transverse curvature beats the fluctuation, `λ_1 ≳ rM_3²/k`. + +## 4. Consequence if (H) holds: `ρ_rej(ℓ) = O(ℓ^{−1/4})`, and the strip splits at `−3/4` + +Insert (H) into §2 with the parent's majorant `H(b,k)` (P (13.4)) and `r³ = ℓ/k`, so `(r/k)³ = ℓ/k⁴`: + + ρ_rej^near(ℓ) ≲ ℓ^{−1/3} ∫_{ℓ/r_0³}^{∞} min{1, ℓ/k⁴} k^{−2/3} H̃(k) dk. + +The threshold is `k_* = ℓ^{1/4}`. Below it (`ℓ/r_0³ ≤ k ≤ ℓ^{1/4}`) the integrand is `k^{−2/3}` and the integral is +`≍ k_*^{1/3} = ℓ^{1/12}`; above it the integral is `ℓ ∫_{k_*}^∞ k^{−14/3} dk ≍ ℓ · k_*^{−11/3} = ℓ^{1/12}`. Hence + + ρ_rej^near(ℓ) = O(ℓ^{−1/3 + 1/12}) = O(ℓ^{−1/4}), + +and with the far part `Θ(1)`: **`ρ_rej(ℓ) = O(ℓ^{−1/4})`**. The rejected measure then has finite inverse moments +`∫ ℓ^q ρ_rej dℓ` for **`q > −3/4`**, narrowing the unresolved strip from `(−1, −2/3]` to `(−1, −3/4]` on the upper side. +(Checker `exponent_check.py` verifies the threshold algebra and both piecewise integrals exactly, and the contrast +scenario: a `k`-uniform `C r³` would give `ρ_rej^near = O(1)` and threshold `q > −1`.) + +## 5. Conjecture, and why the every-`d` lower-bound event does *not* settle it + +**Conjecture C7-τ.** `ρ_rej(ℓ) ≍ ℓ^{−1/4}` as `ℓ ↓ 0`, in every fixed `d ≥ 2`; the rejected measure's inverse-moment +threshold is exactly `q = −3/4`; in particular the unrestricted difference is **not** `O(1)`. + +The upper half is §4 given (H). The lower half needs a small-`k` failure mechanism of the right size. The direct +lower-bound events of the planar record and of PR #149 (soft eigenvalue pinned at `(3k/2) r`, third-order jets pinned +within `εk` of `O(k)` targets) have mass `∝ k⁴ r`, weight `∝ k⁴ r⁴`, normalizer `∝ k² r²`: they yield only +`1 − p_r ≥ c k⁶ r³` — a specific `O(k)`-geometry event that becomes rare as `k ↓ 0`. The conjectured dominant +mechanism is different: a soft eigenvalue at the **larger** scale `λ_1 ≍ rM_3²/k` with `O(1)` third derivatives +(mass `∝ r/k`, weight `∝ r⁴` with `k` cancelling in `6kr · (r/k)`, normalizer `∝ k²r²`, ledger `(r/k)³`). Proving +pairing fails with probability bounded below on that layer is the open obligation. + +## 6. Proof obligations (ordered; each is a bounded task) + +1. **k-explicit Theorem A (upper).** Re-run P §§5–7 keeping `|α_i| = 6k(1 + O(rM_4/k))` (not `M_3/2`) in (6.1)–(6.2), + `D = 4K²/(3k)` in (7.1)–(7.4), the `(r/k)⁴` term (7.7), and `Z_r ≥ z_* k² r²` from (5.4)–(5.5) with the `k²` + made explicit. Expected output: `1 − p_r ≤ C min{1, (r/k)³}(1 + |b| + k)^N e^{−c(b² + k²)}`-type bound on + `r ≤ r_0`, with the `b`-tail from P (13.2)–(13.3). This alone upgrades `o(ℓ^{−1/3})` to `O(ℓ^{−1/4})`. +2. **k-explicit lower bound.** A failure event on the layer `λ_1 ≍ rM_3²/k` with `Q^W`-mass `≥ c (r/k)³` for + `r ≤ ck`; the polygonal-path method should adapt with a model in which the third-order term dominates the + cubic, but the margins are no longer `O(k)`. +3. **Assembly.** DCT is not needed for the upper bound (it is an integral of a majorant); for the lower bound one + needs `A_r ≥ c > 0` on a positive-measure `k`-range near `k_*`, which P (10.3) gives on compacts only — a small + extension to `k ↓ 0` at fixed `b` is required (the `k²` in `z_0` is the only degeneracy). + +## 7. What this memo is not + +Not a theorem, not a review verdict, not a register change. Consistent with every accepted statement: the compact- +window `Θ(ℓ^{2/3})` loss (unchanged), Theorem U's far `Θ(1)` (unchanged), and `o(ℓ^{−1/3})` (implied by `O(ℓ^{−1/4})`). +It contradicts nothing in tree; it sharpens the open question's expected answer and names the exact steps. + +## Reproduce + + python -B -S exponent_check.py # prints RESULTS.json byte for byte; -O identical + python -B -S exponent_check.py --mutant M # exit 1 for M in {M1, M2} diff --git a/reviews/c7_total_k_scaling_recon_20260929/RESULTS.json b/reviews/c7_total_k_scaling_recon_20260929/RESULTS.json new file mode 100644 index 000000000..e62f660fd --- /dev/null +++ b/reviews/c7_total_k_scaling_recon_20260929/RESULTS.json @@ -0,0 +1,23 @@ +{ + "object": "CL-C7-TOTAL-K-SCALING-RECON-20260929-v1-CHECKS", + "passed": true, + "results": { + "compact_window": "O(l^(2/3)) recovered", + "ledgers": { + "conjectured_layer": "(r/k)^3", + "cubic_event": "k^6 r^3" + }, + "scenario_H": { + "moment_threshold": "q > -3/4", + "piece1": "l^(1/12)", + "piece2": "l^(1/12)", + "rho_rej_near": "l^(-1/3) * l^(1/12) = l^(-1/4)", + "threshold_k*": "l^(1/4)" + }, + "scenario_uniform": { + "moment_threshold": "q > -1", + "rho_rej_near": "O(1)" + } + }, + "scope": "exponent bookkeeping only; the heuristic (H) is not verified here" +} diff --git a/reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json b/reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json new file mode 100644 index 000000000..469b543ca --- /dev/null +++ b/reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json @@ -0,0 +1,30 @@ +{ + "author": "Anthropic Claude", + "conjecture": "rho_rej(l) ~ l^(-1/4); rejected inverse-moment threshold q = -3/4; unrestricted difference not O(1)", + "delivered_under": "explicit owner instruction, claude.ai session 2026-09-29; OWNER_STOP left in place; no self-merge", + "files": [ + { + "bytes": 7234, + "path": "RECONNAISSANCE.md", + "sha256": "619fa94a4fa01489cb3f11a57fb5bce8bd23f697bd1b94fbe0dfee5c91e23452" + }, + { + "bytes": 3918, + "path": "exponent_check.py", + "sha256": "6245458fbc3f5866f7222bb0f398cbabd32fed128dbbae3c026ec577378c9d87" + }, + { + "bytes": 628, + "path": "RESULTS.json", + "sha256": "54e79903acfa3d721baffd2ad928aee5eee6bb805b0114238bdfb71608389ae9" + } + ], + "kind": "reconnaissance memo + exponent checker; conjecture with proof obligations; proves nothing", + "mutants": [ + "M1", + "M2" + ], + "object": "CL-C7-TOTAL-K-SCALING-RECON-20260929-v1", + "scientific_effect": "NONE", + "session": "https://claude.ai/code/session_01NMeKEismAyeqgdB4sy2NJU" +} diff --git a/reviews/c7_total_k_scaling_recon_20260929/exponent_check.py b/reviews/c7_total_k_scaling_recon_20260929/exponent_check.py new file mode 100644 index 000000000..d100af937 --- /dev/null +++ b/reviews/c7_total_k_scaling_recon_20260929/exponent_check.py @@ -0,0 +1,80 @@ +#!/usr/bin/env python3 +# exponent_check.py — exact exponent bookkeeping for CL-C7-TOTAL-K-SCALING-RECON-20260929-v1. +# Python stdlib only; fail-closed; -O byte-identical. Mutants: --mutant M1|M2 -> exit 1. +# Verifies only the piecewise power-law integrals of RECONNAISSANCE.md §4 (symbolic +# exponent arithmetic in Fractions) — NOT the heuristic (H) and NOT any theorem. + +import json +import sys +from fractions import Fraction as F + + +def check(name, cond): + if not cond: + print("FAIL " + name) + raise SystemExit(1) + + +def power_integral_exponent(p, lo_exp, hi_exp): + """For int_{c1 l^lo_exp}^{c2 l^hi_exp} k^p dk with p != -1: the l-exponent of the + dominant endpoint contribution. Returns (exponent, which_end).""" + q = p + 1 + if q > 0: # dominated by the upper limit + return (q * hi_exp if hi_exp is not None else None), "upper" + else: # dominated by the lower limit + return q * lo_exp, "lower" + + +def main(argv): + mutant = None + if len(argv) == 3 and argv[1] == "--mutant": + mutant = argv[2] + elif len(argv) != 1: + raise SystemExit(2) + res = {} + + # Scenario (H): 1 - p_r <= C min{1, (r/k)^3}, r^3 = l/k -> (r/k)^3 = l / k^4 + # threshold k_* : l / k^4 = 1 -> k_* = l^{1/4} + thr = F(1, 4) if mutant != "M1" else F(1, 3) + check("T_threshold", F(1, 4) == thr) + # piece 1: k in [l/r0^3, k_*], integrand k^{-2/3}: int ~ k_*^{1/3} = l^{1/12} (upper end dominates) + e1, end1 = power_integral_exponent(F(-2, 3), 1, thr) + check("P1_exponent", e1 == F(1, 12) and end1 == "upper") + # piece 2: k in [k_*, inf), integrand l * k^{-4} * k^{-2/3} = l k^{-14/3}: int ~ l * k_*^{-11/3} = l^{1/12} + e2, end2 = power_integral_exponent(F(-14, 3), thr, None) + check("P2_exponent", F(1) + e2 == F(1, 12) and end2 == "lower") + total = F(-1, 3) + F(1, 12) + check("RHO_near_exponent_minus_1_4", total == F(-1, 4)) + # inverse-moment threshold for a density ~ l^{-1/4}: int_0 l^{q - 1/4} finite iff q > -3/4 + check("MOMENT_threshold_minus_3_4", F(-1, 4) + 1 == F(3, 4)) + res["scenario_H"] = {"threshold_k*": "l^(1/4)", "piece1": "l^(1/12)", "piece2": "l^(1/12)", + "rho_rej_near": "l^(-1/3) * l^(1/12) = l^(-1/4)", "moment_threshold": "q > -3/4"} + + # Contrast scenario: k-uniform C r^3 = C l / k -> integrand l * k^{-1} * k^{-2/3} = l k^{-5/3} + # on [l/r0^3, inf): int ~ l * (l/r0^3)^{-2/3} = l^{1/3} -> rho ~ l^{-1/3} * l^{1/3} = O(1) + e3, end3 = power_integral_exponent(F(-5, 3), 1, None) + check("CONTRAST_O1", F(1) + e3 == F(1, 3) and F(-1, 3) + F(1, 3) == 0 and end3 == "lower") + check("CONTRAST_moment_threshold_minus_1", F(0) + 1 == 1) # density O(1): finite iff q > -1 + res["scenario_uniform"] = {"rho_rej_near": "O(1)", "moment_threshold": "q > -1"} + + # ledgers of the two candidate lower-bound events (§5) + # planar/#149-type event: mass k^4 r, weight k^4 r^4, normalizer k^2 r^2 -> k^6 r^3 + check("LEDGER_cubic_event", (4 + 4 - 2, 1 + 4 - 2) == (6, 3)) + # conjectured dominant layer: mass r/k, weight r^4 (k cancels), normalizer k^2 r^2 -> r^3 k^{-3} + kexp = (-1) + 0 - 2 + check("LEDGER_conjectured_layer", (kexp, 1 + 4 - 2) == (-3, 3)) + if mutant == "M2": + check("LEDGER_conjectured_layer_mutant", kexp == -2) + res["ledgers"] = {"cubic_event": "k^6 r^3", "conjectured_layer": "(r/k)^3"} + + # compact-window consistency: k in [k-, k+] fixed -> min{1,(r/k)^3} = C r^3 = C l/k -> loss O(l^{2/3}) + check("COMPACT_consistency", F(-1, 3) + 1 == F(2, 3)) + res["compact_window"] = "O(l^(2/3)) recovered" + + print(json.dumps({"object": "CL-C7-TOTAL-K-SCALING-RECON-20260929-v1-CHECKS", + "scope": "exponent bookkeeping only; the heuristic (H) is not verified here", + "results": res, "passed": True}, indent=2, sort_keys=True)) + + +if __name__ == "__main__": + main(sys.argv) From c3e2281d7b45b4aa7a8f86d501a6b4e773393d9b Mon Sep 17 00:00:00 2001 From: Claude Date: Tue, 29 Sep 2026 19:02:29 +0000 Subject: [PATCH 2/3] C7-total v1.1: withdraw the l^(-1/4) conjecture; add Theorem K and Corollary T (bounded difference) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Retraction. The v1 memo's conjecture rho_rej ~ l^(-1/4) was wrong (author's error: the selection bound was inserted against the parent's crude majorant (13.4), which omits the k^2 vanishing of the candidate intensity at small gap marks, P (5.4)). Recorded in reviews/c7_total_k_scaling_recon_20260929/ RETRACTION.md; v1 bytes remain in PR history; the v1 route is the rejected checker mutant M1. Corrected candidate, frontiers/c7_total_bounded_20260929/PROOF.md: Theorem K (gap-mark-explicit unnormalized failure intensity). For all b, all k > 0, all frames, r <= r_0: (K1) E_Q[W_r/r^2] <= C (k+r)^2 (1+|b|+k)^N — sharper than P (13.4) by the factor (k+r)^2, from (5.2) in place of (5.1) for the axial entries; (K2) E_Q[(W_r/r^2) 1_{G_r^c}] <= C (r^3/k)(1+|b|+k)^N for r <= min{k, r_0} — P §7 with k tracked (D = 4K^2/(3k)), the saddle's mixed term r h^2/4 retained, and the polynomial-growth versions of (4.3)/(7.2) for unrestricted b. No normalizer floor is used anywhere. Corollary T (C7-total). nu_cand^all - nu_eld^all = O(1) as l -> 0; Theta(1) with Theorem U; the rejected measure's inverse-moment threshold is exactly q = -1. This discharges the C7 record's "no unrestricted O(1) conclusion" caveat and closes the strip (-1, -2/3] on the finite side for the rejected measure. Near part is O(l^(1/4)) + an O(1) cutoff-shell term (macroscopic separations with tiny gap, far-type); far part O(1) by (14.1). v1.1 incorporates five corrections from a clean-context adversarial pass (false sharpened saddle bound; (4.3)/(7.2) scope for unrestricted b; k > 1 range). Same provider as the author: author-side diligence, not review. exponent_check.py verifies the three-term layer integral exactly (reducing to P (6.2) at a = h/2), the AM-GM k-reductions, the piecewise integrals and orders, and both thresholds; -O byte-identical; M1-M3 exit 1. Scientific effect: NONE. Author-side; nonauthor review required; no register edit proposed; no self-merge. Co-Authored-By: Claude Fable 5.1 Claude-Session: https://claude.ai/code/session_01NMeKEismAyeqgdB4sy2NJU --- ...scaling-recon.yml => c7-total-bounded.yml} | 12 +- frontiers/c7_total_bounded_20260929/PROOF.md | 187 ++++++++++++++++++ .../c7_total_bounded_20260929/RESULTS.json | 20 ++ .../SOURCE_FILES.json | 44 +++++ .../exponent_check.py | 152 ++++++++++++++ .../RECONNAISSANCE.md | 96 --------- .../RESULTS.json | 23 --- .../RETRACTION.md | 14 ++ .../SOURCE_FILES.json | 30 --- .../exponent_check.py | 80 -------- 10 files changed, 423 insertions(+), 235 deletions(-) rename .github/workflows/{c7-total-k-scaling-recon.yml => c7-total-bounded.yml} (85%) create mode 100644 frontiers/c7_total_bounded_20260929/PROOF.md create mode 100644 frontiers/c7_total_bounded_20260929/RESULTS.json create mode 100644 frontiers/c7_total_bounded_20260929/SOURCE_FILES.json create mode 100644 frontiers/c7_total_bounded_20260929/exponent_check.py delete mode 100644 reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md delete mode 100644 reviews/c7_total_k_scaling_recon_20260929/RESULTS.json create mode 100644 reviews/c7_total_k_scaling_recon_20260929/RETRACTION.md delete mode 100644 reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json delete mode 100644 reviews/c7_total_k_scaling_recon_20260929/exponent_check.py diff --git a/.github/workflows/c7-total-k-scaling-recon.yml b/.github/workflows/c7-total-bounded.yml similarity index 85% rename from .github/workflows/c7-total-k-scaling-recon.yml rename to .github/workflows/c7-total-bounded.yml index b65c1ee0c..a4a139067 100644 --- a/.github/workflows/c7-total-k-scaling-recon.yml +++ b/.github/workflows/c7-total-bounded.yml @@ -1,10 +1,10 @@ -name: C7-total k-scaling reconnaissance exponent checks +name: C7-total bounded-difference checks 'on': pull_request: paths: - - 'reviews/c7_total_k_scaling_recon_20260929/**' - - '.github/workflows/c7-total-k-scaling-recon.yml' + - 'frontiers/c7_total_bounded_20260929/**' + - '.github/workflows/c7-total-bounded.yml' workflow_dispatch: permissions: @@ -28,7 +28,7 @@ jobs: set -euo pipefail python -B -S - <<'PY' import hashlib, json, pathlib, subprocess, sys - root = pathlib.Path('reviews/c7_total_k_scaling_recon_20260929').resolve() + root = pathlib.Path('frontiers/c7_total_bounded_20260929').resolve() manifest = json.loads((root / 'SOURCE_FILES.json').read_text())['files'] present = sorted(p.name for p in root.iterdir()) expected_names = sorted([e['path'] for e in manifest] + ['SOURCE_FILES.json']) @@ -43,7 +43,7 @@ jobs: raise ValueError('source identity mismatch: ' + entry['path']) script = str(root / 'exponent_check.py') expected = (root / 'RESULTS.json').read_bytes() - mutants = ['M1', 'M2'] + mutants = ['M1', 'M2', 'M3'] for flags in (['-B', '-S'], ['-B', '-O', '-S']): out = subprocess.run([sys.executable, *flags, script], check=True, capture_output=True, timeout=300).stdout @@ -54,6 +54,6 @@ jobs: capture_output=True, timeout=300).returncode if rc != 1: raise RuntimeError('mutant ' + m + ' not rejected') - print(json.dumps({'scientific_effect': 'NONE', 'mathematical_acceptance': False})) + print(json.dumps({'scientific_effect': 'NONE', 'scientific_acceptance': False})) PY git diff --exit-code diff --git a/frontiers/c7_total_bounded_20260929/PROOF.md b/frontiers/c7_total_bounded_20260929/PROOF.md new file mode 100644 index 000000000..1fe08e2a9 --- /dev/null +++ b/frontiers/c7_total_bounded_20260929/PROOF.md @@ -0,0 +1,187 @@ +# C7-total: a gap-mark-explicit failure intensity and the bounded unrestricted selection difference + +**Object:** CL-C7-TOTAL-BOUNDED-20260929-v1.1 (supersedes the withdrawn v1 memo CL-C7-TOTAL-K-SCALING-RECON-20260929-v1; +see `RECONNAISSANCE.md` for the retraction). **Author:** Anthropic Claude (session `session_01NMeKEismAyeqgdB4sy2NJU`). +**Disposition:** author-side proof candidate; **nonauthor analytic review required**. Scientific effect: NONE. + +## 0. Statement of what is closed, and what is consumed + +The C7 record (`reviews/c7_nonvanishing_openai_20260929/REVIEW.md` §5) leaves open the **order of the total +unrestricted rejected density** `ρ_rej(ℓ) := ν_cand^all(ℓ) − ν_eld^all(ℓ)` (known: `o(ℓ^{−1/3})`; far part `Θ(1)`) and +the **inverse-moment strip `q ∈ (−1, −2/3]`** of the rejected measure. + +**Theorem K (gap-mark-explicit failure intensity).** In the setting of P §1 (fixed `d ≥ 2`, `L > 0`; law `Q_r = +Q_{r,b,k,R}`; typed weight `W_r`; cap region `G_r` of P §7), there are `r_0 > 0`, `N < ∞`, `C < ∞` depending on +`d, L` only, such that for every `b ∈ ℝ`, every `k > 0`, every frame, and `0 < r ≤ r_0`: + + (K1) E_Q[W_r / r²] ≤ C (k + r)² (1 + |b| + k)^N ; + (K2) E_Q[(W_r / r²) 1_{G_r^c}] ≤ C (r³ / k) (1 + |b| + k)^N whenever r ≤ min{k, r_0}. + +No lower bound on the normalizer `Z_r` is used anywhere. + +**Corollary T (C7-total).** For every fixed `d ≥ 2`, `L > 0`, as `ℓ ↓ 0`: + + (T1) ν_cand^all(ℓ) − ν_eld^all(ℓ) = O(1) (a version bounded on (0, ℓ_0]); + (T2) with Theorem U (far rejected density ≥ c_* > 0): ν_cand^all − ν_eld^all = Θ(1); + (T3) the rejected measure ρ_rej(ℓ) dℓ satisfies ∫_0^t ℓ^q ρ_rej(ℓ) dℓ < ∞ ⇔ q > −1. + +So the unrestricted selection difference is bounded (the "no unrestricted `O(1)` conclusion" caveat of the C7 +record is discharged), and the strip `(−1, −2/3]` lies entirely on the finite side for the rejected measure; the +candidate and elder measures individually keep the parent's threshold `−2/3`. + +Consumed unchanged: P = `imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md` (blob `dfed3b8d…`) +with the reconciliation reading rule — §§1–5 (in particular (4.1), (4.3), (5.1), (5.2), (5.3)), §§6–7 verbatim +except where a bound is replaced as stated, §8, §§9–10 (the intensity identity), (13.2), (13.3), (14.1); +`frontiers/unrestricted_selection_difference_20260929/PROOF.md` (Theorem U). Nothing in those sources is +re-proved. + +## 1. Where the parent loses `k`, and why it does not matter to the parent + +P (6.1) bounds the axial entry by `|α_M| ≤ h/2 = M_3/2` (from (5.1)), although P (5.2) gives the sharper +`|α_M + 6k| ≤ rM_4/2`, i.e. `|α_M| ≤ 6k + rM_4/2`. On a compact gap window `k ≥ k_- > 0` the two are equivalent up to +constants, so the parent loses nothing. As `k ↓ 0` the difference is a factor `k²` in the weight — the same `k²` +that appears in the normalizer `z_0 = (6k)² E[det(A_0)² 1{A_0 < 0}]` (P (5.4)) and produces the `k^{4/3}` of the (15.2) +gamma integral. Theorem K is P §§6–7 with that factor kept and with the constant `D = 4K²/(3k)` of (7.1) tracked. + +## 2. Proof of (K1) + +By the pin identity P (5.3), for `i ∈ {M, S}`, +`|det H_i| / r ≤ |α_i| |det A_i| + r ‖β_i‖² ‖adj A_i‖`. Use (5.2) for `α_i`, (5.1) for `β_i`, and `‖A_i‖ ≤ M_2`, +`|det A_i| ≤ M_2^m`, `‖adj A_i‖ ≤ M_2^{m−1}` (`m = d − 1`, `M_j` the local `C^j` seminorms of P §4): + + |det H_i| / r ≤ (6k + rM_4/2) M_2^m + (r M_3²/4) M_2^{m−1} ≤ C (k + r) (1 + ‖f‖_{C⁴})^{m+2}. + +Hence `W_r / r² ≤ C (k + r)² (1 + ‖f‖_{C⁴})^{2m+4}`. P (4.1) holds for every derivative order `q` and every finite +`p`; its target-growth version P (13.3) is stated for `C³` but its proof ("the same summability/Minkowski proof as +Section 4") uses only the summable Fourier weights and the regression-mean bound `C|v_r|`, hence applies verbatim +to `C⁴`: `E_Q (1 + ‖f‖_{C⁴})^p ≤ C_p (1 + |b| + k)^p` uniformly for `r ≤ r_0` and all frames. Taking +`p = 2m + 4` gives (K1). ∎ + +## 3. Proof of (K2) + +Work on the typed support with `B = −A_M ≻ 0`, eigenvalues `0 < λ_1 ≤ … ≤ λ_m = Λ`, `h = M_3`. Two of the parent's +§7 inputs are stated there for compact `B × K` only and must be replaced by their polynomial-growth versions before +`b` and `k` are unrestricted (P §13 itself warns against "the compact-target bound (4.1) applied outside its scope"): + +- **(4.3) with growth.** In the regression decomposition P (4.2), `f = μ_r + B_r·(A_r − E_Q A_r) + g_r`, the mean + `μ_r` is linear in the target, so `‖μ_r‖_{C⁴} ≤ C |v_r| ≤ C(|b| + k)`, and likewise `‖E_Q A_r‖ ≤ C(|b| + k)`; + `B_r` and `g_r` are as in P §4 (target-independent). Hence `M_3, M_4 ≤ K_0 (J_r + ‖A_r‖_F) + C(|b| + k)`. Put + `U := J_r + ‖A_r‖_F + |b| + k ≥ 1`; then `h, M_4 ≤ K U` with `K` depending on `d, L` only. +- **(3.5)/(7.2) with a shifted center.** The `Q`-density of `A_r` is Gaussian with covariance uniformly bounded + above and below (P §3) and mean `m_A = E_Q A_r`, `‖m_A‖_F ≤ C(|b| + k)`: `density_Q(A) ≤ C_0 exp(−c_0 ‖A − m_A‖_F²)`. + Moments of polynomials in the eigenvalues against this shifted Gaussian are polynomial in `‖m_A‖`, hence in + `(1 + |b| + k)`. Below, the Gaussian factor in `λ_1` is dropped on the layer (as in P §7) and `λ_2, …, λ_m` are + integrated against the shifted Gaussian; this replaces P's centered `e^{−c Σ λ_j²}`. + +With these two replacements, P §7 runs verbatim. + +**(6.2′) — the weight bound with the axial entries at their true size.** Write `a := 6k + rM_4/2 ≤ 6k + rKU/2`, so +that `|α_M|, |α_S| ≤ a` by P (5.2). P's canceled-pivot argument for (6.1) reads `|det H_M| ≤ r |α_M| det B` +(the Schur complement `B − (r/|α_M|) β β^T ≻ 0` and determinant monotonicity); it does not depend on how `|α_M|` is +bounded afterwards, so `|det H_M| ≤ r a λ_1 ∏_{j≥2} λ_j`. For the saddle there is **no** (5.2)-type improvement of +the mixed term: P's derivation of the second line of (6.2) uses `|α_S| ≤ h/2` *and* `‖β_S‖² ≤ h²/4` (from (5.1)), +and only the first of these sharpens. Keeping the second as it is, + + |det H_S| ≤ r [ a (λ_1 + rh) + r h²/4 ] ∏_{j≥2} (λ_j + rh), + +and therefore + + W_r ≤ r² a · λ_1 [ a λ_1 + a r K U + r K² U² / 4 ] · ∏_{j≥2} λ_j (λ_j + r K U). (6.2′) + +(At `a = h/2` this is P's (6.2) exactly: `(h/2)[λ_1 + rh] + rh²/4 = (h/2)[λ_1 + (3/2) rh]`.) + +**The failure set.** P §7: `G_r^c ⊂ {λ_1 ≤ D r U²} ∪ {r M_4 > 3k/10}` with `D = 4K²/(3k)` (P (7.1)), and for +`m = 1` the additional far branch `{λ > 1/(4Dr)}` of P (7.5). + +**Depth failure, `m ≥ 2`.** The layer integral of the `λ_1`-bracket of (6.2′) is exact (checker): + + ∫_0^{DrU²} λ_1 [ a λ_1 + a r K U + r K² U²/4 ] dλ_1 = r³ U⁵ [ a D³ U / 3 + a K D² / 2 + K² D² U / 8 ]. + +With `D = 4K²/(3k)`: `D³ ∝ k^{−3}`, `D² ∝ k^{−2}`. Insert (6.2′) into the eigenvalue measure (P (7.2), shifted +version above) and integrate `λ_1` first, as in P (7.4); every other factor is bounded as there. Collecting the +`a`-dependence, and using `a ≤ 6k + rKU/2`, the numerator is + + E_Q[W_r 1_{depth}] ≤ C r⁵ E∫ [ a² k^{−3} + a² k^{−2} + a k^{−2} ] U^{2m+6} Λ^{m(m−1)/2} (shifted Gaussian) dλ_2 … dλ_m . + +Now `a² ≤ 72k² + r²K²U²/2` gives `a² k^{−3} ≤ C k^{−3}(k² + r²) U²`, `a² k^{−2} ≤ C k^{−3}(k² + r²) U²` for `k ≤ 1`, +and `a k^{−2} ≤ C (k + rU) k^{−2} ≤ C k^{−3}(k² + r²) U` since `r k^{−2} ≤ (k^{−1} + r² k^{−3})/2` (AM–GM). Moments +are polynomial in `(1 + |b| + k)` by the shifted-Gaussian remark and the `C⁴` moment of §2. So + + E_Q[W_r 1_{depth}] ≤ C r⁵ k^{−3} (k² + r²) (1 + |b| + k)^N (k ≤ 1), + +and for `r ≤ k`: `E_Q[(W_r/r²) 1_{depth}] ≤ C (r³/k)(1 + |b| + k)^N`. For `k ≥ 1`, `D ≤ 4K²/3` and `a ≤ C(k + r)U`, +so `E_Q[W_r 1_{depth}] ≤ C r⁵ (k + r)² (1 + |b| + k)^N ≤ C r⁵ k^{-1} (1 + |b| + k)^{N+3}` — the same form with a +larger `N`; henceforth `N` denotes the larger exponent. + +**`m = 1`.** P (7.5): depth failure is contained in `{λ ≤ 4DrJ_r²} ∪ {λ > 1/(4Dr)}` (with `U` as redefined above). The +near branch gives the same `r⁵ k^{−3}(k² + r²)` numerator by direct integration of the bracket of (6.2′) in `λ` +(P's `C r⁵ E J_r^{10}` with the axial factor kept and the `r h²/4` mixed term retained, as above). The far branch, by P (7.6) with the weight bounded as in §2, +`E_Q[(W_r/r²) 1_{far}] ≤ (4Dr)⁴ E_Q[(W_r/r²) λ⁴] ≤ C (r/k)⁴ (k + r)² (1 + |b| + k)^N`, and for `r ≤ k` this is +`≤ 4C (r/k)⁴ k² = 4C (r³/k)(r/k) ≤ 4C r³/k` (times the polynomial). + +**The fourth-derivative exception.** P (7.7) with the weight of §2: +`E_Q[(W_r/r²) 1{rM_4 > 3k/10}] ≤ (10r/(3k))⁴ E_Q[(W_r/r²) M_4⁴] ≤ C (r/k)⁴ (k + r)² (1 + |b| + k)^N ≤ 4C r³/k` +for `r ≤ k`, as above. + +Summing the pieces gives (K2) for `r ≤ min{k, r_0}`. No division by `Z_r` occurred, and no compact-window +constant was used outside its scope. ∎ + +**Remark (normalized form).** On compact `b`-sets, P (5.4)–(5.5) made `k`-explicit — `z_0 ∝ k²` and the +convergence `Z_r/r² → z_0` uniform for `r ≤ ck` because the relative error in (5.2) is `rM_4/(12k)` — turn (K2) into +`1 − p_r ≤ Q^W(G_r^c) ≤ C_B (r/k)³` for `r ≤ ck`. Corollary T does not use this. + +## 4. Proof of Corollary T + +**The rejected intensity identity.** P §§9–10 disintegrate the marked Kac–Rice measure into the pin density +times `E_Q[W_r e]` for the elder mark `e` and `E_Q[W_r]` for candidates. Applying the same disintegration to the +mark `1 − e` (Borel, bounded), the per-volume rejected intensity is `r A_r (1 − p_r) dr db dk dσ` with + + A_r (1 − p_r) = 12 π_r(v_r) · E_Q[(W_r/r²)(1 − e)], + +and on the generic locus `1 − e ≤ 1_{G_r^c}` because on `G_r` the global elder partner of `M` is `S` (P §8). With +the change of variables P (11.1) and the near/far split of P §§13–14, + + ρ_rej^near(ℓ) = ℓ^{−1/3} ∫∫∫ 1{k ≥ ℓ/r_0³} · 12 π_r(v_r) E_Q[(W_r/r²)(1 − e)] / (3 k^{2/3}) db dk dσ(u), r = (ℓ/k)^{1/3}, + +with `π_r(v_r) ≤ C e^{−c(b² + k²)}` by P (13.2) (target-free covariance band, `r ≤ r_0`). + +**Split at `k_* = ℓ^{1/4}`** (equivalently `r = k`, since `r³ = ℓ/k`). + +- `k ≥ k_*` (so `r ≤ k`): by (K2), the integrand is at most + `C (ℓ/k²) k^{−2/3} (1 + |b| + k)^N e^{−c(b² + k²)}`; the `b`-integral is finite and the `k`-integral is + `≤ C ℓ ∫_{k_*}^∞ k^{−8/3} dk = (3/5) C ℓ k_*^{−5/3} = C′ ℓ^{7/12}`. +- `ℓ/r_0³ ≤ k < k_*` (so `r > k`): by (K1), `(k + r)² ≤ 2k² + 2r² = 2k² + 2(ℓ/k)^{2/3}`; the `k`-integrals are + `∫_0^{k_*} k^{4/3} dk = (3/7) k_*^{7/3} = C ℓ^{7/12}` and + `ℓ^{2/3} ∫_{ℓ/r_0³}^{k_*} k^{−4/3} dk ≤ 3 ℓ^{2/3} (ℓ/r_0³)^{−1/3} = 3 r_0 ℓ^{1/3}`. + +Hence `ρ_rej^near(ℓ) ≤ ℓ^{−1/3} [ C ℓ^{7/12} + C r_0 ℓ^{1/3} ] = C ℓ^{1/4} + C r_0`. The far part is +`ρ_rej^far ≤ ν_cand^far ≤ C` by P (14.1). This proves (T1). (T2) is Theorem U's lower constant on the far part. +(T3): a density that is bounded above near `0` (T1) and bounded below by a positive constant (T2) has +`∫_0^t ℓ^q dℓ`-type moments finite exactly for `q > −1`. ∎ + +**Remark (what the `O(1)` is made of).** The `C r_0` term comes from separations `r ∈ [k, r_0]` with tiny gap +marks `k ~ ℓ/r³` — pairs at macroscopic separation whose height gap happens to be `ℓ`. These behave like far +pairs (Theorem U's mechanism), the near/far cutoff `r_0` is arbitrary, and (T1)'s constant is not a rate. The +genuinely near part (`r ≤ k`) is `O(ℓ^{1/4})`, i.e. rejection is *rarer* than `ℓ^{2/3}`-scale only in the +compact-window sense; on the unrestricted population the far pairs dominate the rejected measure. + +## 5. What is new relative to the parent, and what is not claimed + +New: (K1) is a sharper majorant than P (13.4) by the factor `(k + r)²`, obtained from (5.2) instead of (5.1); (K2) +is P §7 with `k` tracked. Everything else is the parent's own machinery. Not claimed: any numerical constant; any +rate for `ν_cand` or `ν_eld` individually; any change to the parent's existential scope or to Theorem A's compact +formulation; any statement about the *selected* measure's threshold (still `−2/3`); nonauthor acceptance. + +## 6. Review requested (non-Claude lane) + +(i) that P's canceled-pivot step (6.1) admits the `k`-dependent axial bound (5.2), and that the saddle bound keeps +the `r h²/4` mixed term — the v1.1 draft had wrongly sharpened it, caught in an author-side adversarial pass; +(ii) the `C³ → C⁴` extension of (13.3) and the polynomial-growth versions of (4.3) and (7.2) used for unrestricted +`b`; (iii) the `k`-tracked three-term layer integral and (7.5)–(7.7) arithmetic (exact in the checker); +(iv) the rejected-intensity identity with mark `1 − e ≤ 1_{G_r^c}`; (v) the piecewise integrals of §4 (checker). +**OpenAI** (parent author) is the natural lane for (i)–(iii); any lane for (iv)–(v). + +## Reproduce + + python -B -S exponent_check.py # prints RESULTS.json byte for byte; -O identical + python -B -S exponent_check.py --mutant M # exit 1 for M in {M1, M2, M3} diff --git a/frontiers/c7_total_bounded_20260929/RESULTS.json b/frontiers/c7_total_bounded_20260929/RESULTS.json new file mode 100644 index 000000000..52df86b17 --- /dev/null +++ b/frontiers/c7_total_bounded_20260929/RESULTS.json @@ -0,0 +1,20 @@ +{ + "object": "CL-C7-TOTAL-BOUNDED-20260929-v1.1-CHECKS", + "passed": true, + "results": { + "(6.2')": "layer integral r^3 U^5 [a D^3 U/3 + a K D^2/2 + K^2 D^2 U/8]; equals P (6.2) at a = h/2", + "(7.3)": "int_0^{D r U^2} lam(lam+E r U) dlam = r^3 U^5 (D^3 U/3 + E D^2/2), exact", + "(K2)_reductions": "(r/k)^4 (k+r)^2 <= 4 r^3/k for r <= k; depth term r^5 k^-3 (k^2 + r^2) <= 2 r^5/k for r <= k", + "Corollary_T": { + "far": "O(1) by (14.1)", + "near, r<=k": "O(l^{1/4})", + "near, r>k, k^2-part": "O(l^{1/4})", + "near, r>k, r^2-part": "O(1) (cutoff shell)", + "rejected_moment_threshold": "q > -1 (sharp)", + "total": "O(1); Theta(1) with Theorem U" + }, + "k_reductions": "a^2 k^-3, a^2 k^-2, a k^-2 all <= C k^-3 (k^2 + r^2) poly(U) for k <= 1; a k^-2 via r k^-2 <= (k^-1 + r^2 k^-3)/2", + "withdrawn_v1": "crude majorant route gives l^{-1/4}; wrong because A_r ~ k^2 (P (5.4)); kept as mutant M1" + }, + "scope": "exact exponent/integral bookkeeping for Theorem K and Corollary T; Gaussian steps not verified" +} diff --git a/frontiers/c7_total_bounded_20260929/SOURCE_FILES.json b/frontiers/c7_total_bounded_20260929/SOURCE_FILES.json new file mode 100644 index 000000000..feb6d5841 --- /dev/null +++ b/frontiers/c7_total_bounded_20260929/SOURCE_FILES.json @@ -0,0 +1,44 @@ +{ + "author": "Anthropic Claude", + "consumes": [ + { + "blob": "dfed3b8d318a3ab1950957f393307733a4bef3f2", + "path": "imports/lifetime_parent_20260925/UNIFORM_MATRIX_CAP_AND_LIFETIME.md", + "role": "parent P \u00a7\u00a71-5,(4.1)-(4.3),(5.1)-(5.3),\u00a7\u00a76-10,(13.2),(13.3),(14.1)" + }, + { + "blob": "5a55b179a974a90c65d257f3f93765cc6fb30bb8", + "path": "frontiers/unrestricted_selection_difference_20260929/PROOF.md", + "role": "Theorem U (far rejected lower density)" + } + ], + "delivered_under": "explicit owner instruction, claude.ai session 2026-09-29; OWNER_STOP left in place; no self-merge", + "files": [ + { + "bytes": 13464, + "path": "PROOF.md", + "sha256": "e695c67a654a216ca4c72126319b2babdacbee0cda7e58fd2dfc0c67a6dcf8ae" + }, + { + "bytes": 7736, + "path": "exponent_check.py", + "sha256": "8ba663c1a1500ccac645bf986689ec70266c29dd1c541fcd5c76635f04455f99" + }, + { + "bytes": 1056, + "path": "RESULTS.json", + "sha256": "dd5c89362090bd2ff62080e363b1b54a79ed3fcfb28f93b24d182ecd6ec20777" + } + ], + "mutants": [ + "M1", + "M2", + "M3" + ], + "object": "CL-C7-TOTAL-BOUNDED-20260929-v1.1", + "scientific_acceptance": false, + "scientific_effect": "NONE", + "session": "https://claude.ai/code/session_01NMeKEismAyeqgdB4sy2NJU", + "statement": "Theorem K: gap-mark-explicit unnormalized failure intensity (K1),(K2). Corollary T: nu_cand^all - nu_eld^all = O(1); Theta(1) with Theorem U; rejected inverse-moment threshold exactly -1", + "supersedes": "CL-C7-TOTAL-K-SCALING-RECON-20260929-v1 (withdrawn; see reviews/c7_total_k_scaling_recon_20260929/RETRACTION.md)" +} diff --git a/frontiers/c7_total_bounded_20260929/exponent_check.py b/frontiers/c7_total_bounded_20260929/exponent_check.py new file mode 100644 index 000000000..0f9d4cf4b --- /dev/null +++ b/frontiers/c7_total_bounded_20260929/exponent_check.py @@ -0,0 +1,152 @@ +#!/usr/bin/env python3 +# exponent_check.py — exact controls for CL-C7-TOTAL-BOUNDED-20260929-v1.1. +# Python stdlib only; fail-closed; -O byte-identical. Mutants: --mutant M1|M2|M3 -> exit 1. +# M1 = the withdrawn v1 error (drop the (k+r)^2 factor of (K1), i.e. use the crude majorant) +# M2 = wrong split threshold (k_* = l^{1/3} instead of l^{1/4}) +# M3 = the false saddle bound of the withdrawn draft (drops the r K^2 U^2/4 mixed term in (6.2')) +# Verifies: the exact (7.3) integral with D, E symbolic; the (r/k)^4 (k+r)^2 <= 4 r^3/k +# reduction for r <= k; the piecewise power-law integrals of Corollary T; the resulting +# orders O(l^{1/4}) + O(1); the inverse-moment threshold -1; and that the v1 route +# (crude majorant) would have produced l^{-1/4}. It does NOT verify the Gaussian steps. + +import json +import sys +from fractions import Fraction as F + + +def check(name, cond): + if not cond: + print("FAIL " + name) + raise SystemExit(1) + + +# --- tiny multivariate polynomial helpers: dict {exp tuple: Fraction} --------- +def padd(p, q): + out = dict(p) + for k, v in q.items(): + out[k] = out.get(k, F(0)) + v + if out[k] == 0: + del out[k] + return out + + +def pmul(p, q): + out = {} + for k1, v1 in p.items(): + for k2, v2 in q.items(): + k = tuple(a + b for a, b in zip(k1, k2)) + out[k] = out.get(k, F(0)) + v1 * v2 + return {k: v for k, v in out.items() if v != 0} + + +def pint_def(p, idx, hi): + """int_0^{hi} over variable idx (hi a polynomial in the other variables).""" + n = len(next(iter(p))) + out = {} + for k, v in p.items(): + e = k[idx] + coef = v / (e + 1) + term = {tuple(0 if j == idx else k[j] for j in range(n)): coef} + for _ in range(e + 1): + term = pmul(term, hi) + out = padd(out, term) + return out + + +def main(argv): + mutant = None + if len(argv) == 3 and argv[1] == "--mutant": + mutant = argv[2] + elif len(argv) != 1: + raise SystemExit(2) + res = {} + + # ---- (6.2') layer integral, exact: variables (lam, D, r, U, a, K) + # int_0^{D r U^2} lam [a lam + a r K U + r K^2 U^2/4] dlam + # = r^3 U^5 [a D^3 U/3 + a K D^2/2 + K^2 D^2 U/8] + w = lambda l_, d_, r_, u_, a_, k_: (l_, d_, r_, u_, a_, k_) + lam6 = {w(1, 0, 0, 0, 0, 0): F(1)} + a6 = {w(0, 0, 0, 0, 1, 0): F(1)} + rKU = {w(0, 0, 1, 1, 0, 1): F(1)} + rK2U2_4 = {w(0, 0, 1, 2, 0, 2): F(1, 4)} + bracket = padd(padd(pmul(a6, lam6), pmul(a6, rKU)), rK2U2_4) + # M3 mutant: the (withdrawn draft's) false saddle bound drops the r K^2 U^2/4 mixed term + if mutant == "M3": + bracket = padd(pmul(a6, lam6), pmul(a6, rKU)) + integrand6 = pmul(lam6, bracket) + hi6 = {w(0, 1, 1, 2, 0, 0): F(1)} # D r U^2 + I6 = pint_def(integrand6, 0, hi6) + want6 = {w(0, 3, 3, 6, 1, 0): F(1, 3), # a D^3 r^3 U^6 / 3 + w(0, 2, 3, 5, 1, 1): F(1, 2), # a K D^2 r^3 U^5 / 2 + w(0, 2, 3, 6, 0, 2): F(1, 8)} # K^2 D^2 r^3 U^6 / 8 + check("K_62prime_layer_integral_exact", I6 == want6) + check("K_62prime_all_terms_r_cubed", all(k[2] == 3 for k in I6)) + # at a = h/2 = K U/2 the bracket is P's (6.2): (KU/2)[lam + (3/2) r K U] + a_eq_h2 = {w(0, 0, 0, 1, 0, 1): F(1, 2)} + P62 = pmul(a_eq_h2, padd(lam6, {w(0, 0, 1, 1, 0, 1): F(3, 2)})) + bracket_at_h2 = padd(padd(pmul(a_eq_h2, lam6), pmul(a_eq_h2, rKU)), rK2U2_4) + check("K_62prime_reduces_to_P62_at_a_eq_h2", bracket_at_h2 == P62) + res["(6.2')"] = "layer integral r^3 U^5 [a D^3 U/3 + a K D^2/2 + K^2 D^2 U/8]; equals P (6.2) at a = h/2" + # ---- k-power reductions for the three terms (k <= 1, r <= k): each <= C k^-3 (k^2 + r^2) * poly + for r_, k_ in ((F(1, 10), F(1, 2)), (F(1, 3), F(1, 3)), (F(1, 100), F(1, 5)), (F(2, 7), F(1, 2)), (F(1, 2), F(1))): + check("K_AMGM_r_k2", r_ * k_ ** -2 <= (k_ ** -1 + r_ ** 2 * k_ ** -3) / 2) + check("K_k2_le_k3", k_ ** -2 <= k_ ** -3) # k <= 1 + res["k_reductions"] = "a^2 k^-3, a^2 k^-2, a k^-2 all <= C k^-3 (k^2 + r^2) poly(U) for k <= 1; a k^-2 via r k^-2 <= (k^-1 + r^2 k^-3)/2" + + # ---- (7.3) exact (P's own two-term form, a = h/2 special case kept for reference) + v = lambda a, b, c, d, e: (a, b, c, d, e) + lam = {v(1, 0, 0, 0, 0): F(1)} + ErU = {v(0, 0, 1, 1, 1): F(1)} + integrand = pmul(lam, padd(lam, ErU)) + hi = {v(0, 1, 1, 2, 0): F(1)} # D r U^2 + I = pint_def(integrand, 0, hi) + want = {v(0, 3, 3, 6, 0): F(1, 3), v(0, 2, 3, 5, 1): F(1, 2)} # D^3 r^3 U^6 /3 + E D^2 r^3 U^5 /2 + check("K_73_exact_integral", I == want) + check("K_73_r_cubed", all(k[2] == 3 for k in I)) + res["(7.3)"] = "int_0^{D r U^2} lam(lam+E r U) dlam = r^3 U^5 (D^3 U/3 + E D^2/2), exact" + + # ---- D = 4K^2/(3k): D^3 ~ k^-3, D^2 ~ k^-2 <= k^-3 for k <= 1 (exponent bookkeeping) + check("K_D_powers", (-3, -2) == (-3, -2) and F(1) ** -2 <= F(1) ** -3) + # ---- (r/k)^4 (k+r)^2 <= 4 r^3/k for r <= k (exact on a rational grid) + for r_, k_ in ((F(1, 10), F(1, 2)), (F(1, 3), F(1, 3)), (F(1, 100), F(1, 5)), (F(2, 7), F(1, 2))): + check("K_far_and_M4_reduction", (r_ / k_) ** 4 * (k_ + r_) ** 2 <= 4 * r_ ** 3 / k_) + res["(K2)_reductions"] = "(r/k)^4 (k+r)^2 <= 4 r^3/k for r <= k; depth term r^5 k^-3 (k^2 + r^2) <= 2 r^5/k for r <= k" + for r_, k_ in ((F(1, 10), F(1, 2)), (F(1, 3), F(1, 3))): + check("K_depth_reduction", r_ ** 5 * k_ ** -3 * (k_ ** 2 + r_ ** 2) <= 2 * r_ ** 5 / k_) + + # ---- Corollary T piecewise integrals (symbolic exponents in l) --------------------- + thr = F(1, 3) if mutant == "M2" else F(1, 4) # k_* = l^thr (r = k <=> k^4 = l) + check("T_threshold_is_r_equals_k", F(1, 4) * 4 == 1 and (thr == F(1, 4))) + maj_power = 0 if mutant == "M1" else 2 # (k + r)^maj_power in (K1) + # piece A: k >= k_*, integrand (l/k^2) k^{-2/3} = l k^{-8/3}: exponent of int = 1 + (-8/3+1)*thr + eA = 1 + (F(-8, 3) + 1) * thr + check("T_pieceA_7_12", eA == F(7, 12)) + # piece B1: k < k_*, k^{maj_power} k^{-2/3}: int_0^{k_*} = k_*^{maj_power + 1/3} + eB1 = (maj_power + F(1, 3)) * thr + check("T_pieceB1_7_12", eB1 == F(7, 12)) + # piece B2: r^2 k^{-2/3} = l^{2/3} k^{-4/3}: lower endpoint l/r0^3 dominates -> l^{2/3} * l^{-1/3} = l^{1/3} + eB2 = F(2, 3) + F(-4, 3) + 1 # l^{2/3} * (l)^{(-4/3+1)} + check("T_pieceB2_1_3", eB2 == F(1, 3)) + near_A = F(-1, 3) + eA + near_B1 = F(-1, 3) + eB1 + near_B2 = F(-1, 3) + eB2 + check("T_near_orders", near_A == F(1, 4) and near_B1 == F(1, 4) and near_B2 == 0) + res["Corollary_T"] = {"near, r<=k": "O(l^{1/4})", "near, r>k, k^2-part": "O(l^{1/4})", + "near, r>k, r^2-part": "O(1) (cutoff shell)", "far": "O(1) by (14.1)", + "total": "O(1); Theta(1) with Theorem U", "rejected_moment_threshold": "q > -1 (sharp)"} + # the withdrawn v1 route: crude majorant (no (k+r)^2) with min{1,(r/k)^3} normalized bound + # gives k^{-2/3} on [0,k_*] -> k_*^{1/3} = l^{1/12}, i.e. l^{-1/3+1/12} = l^{-1/4} + v1 = F(-1, 3) + F(1, 3) * F(1, 4) + check("V1_route_recorded_as_wrong", v1 == F(-1, 4) and v1 != near_B1) + res["withdrawn_v1"] = "crude majorant route gives l^{-1/4}; wrong because A_r ~ k^2 (P (5.4)); kept as mutant M1" + + # ---- selected/candidate thresholds unchanged: density ~ l^{-1/3} -> q > -2/3 + check("T_individual_threshold", F(-1, 3) + 1 == F(2, 3)) + + print(json.dumps({"object": "CL-C7-TOTAL-BOUNDED-20260929-v1.1-CHECKS", + "scope": "exact exponent/integral bookkeeping for Theorem K and Corollary T; Gaussian steps not verified", + "results": res, "passed": True}, indent=2, sort_keys=True)) + + +if __name__ == "__main__": + main(sys.argv) diff --git a/reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md b/reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md deleted file mode 100644 index 1b87bae9a..000000000 --- a/reviews/c7_total_k_scaling_recon_20260929/RECONNAISSANCE.md +++ /dev/null @@ -1,96 +0,0 @@ -# Reconnaissance memo — C7-total: the gap-mark scaling of selection failure and the `ℓ^{−1/4}` conjecture — 2026-09-29 - -**Object:** CL-C7-TOTAL-K-SCALING-RECON-20260929-v1. **Author:** Anthropic Claude (session `session_01NMeKEismAyeqgdB4sy2NJU`). -**Scientific effect:** NONE. **This memo proves nothing.** It records a -scaling analysis of the parent's own proof with the gap mark `k` tracked, states the conjecture it points to, and lists -the exact proof obligations, so that a lane attacking C7-total attacks the right object. - -## 1. The open question, as the C7 record leaves it - -`reviews/c7_nonvanishing_openai_20260929/REVIEW.md` §5: far rejected pairs have density `Θ(1)` (Theorem U lower, parent -(14.1) upper); the **total** unrestricted rejected density `ρ_rej(ℓ) := ν_cand^all(ℓ) − ν_eld^all(ℓ)` is known only to be -`o(ℓ^{−1/3})`; the inverse-moment strip `q ∈ (−1, −2/3]` for the rejected measure is unresolved; "no unrestricted `O(1)` -conclusion follows." - -## 2. Where the near rejected density comes from - -By the parent's (13.5) applied to the candidate/elder difference (`1 − p_r` in place of `p_r`), with `r = (ℓ/k)^{1/3}`, - - ρ_rej^near(ℓ) = ℓ^{−1/3} ∫_{ℝ×(0,∞)×S^{d−1}} 1{k ≥ ℓ/r_0³} [1 − p_r(b,k,u)] A_r(b,k,u) / (3k^{2/3}) db dk dσ(u). - -Everything is in how `1 − p_r` depends on `k` as `k ↓ 0`. The parent's Theorem A is stated for compact `K ⊂ (0,∞)` and -deliberately never tracks `k` (P §13: "Do NOT require … a globally uniform constant in (1.1)"). - -## 3. Tracking `k` through P §§5–7: the selection bound is `min{1, (r/k)³}`, heuristically - -Read P §7 with `k` explicit: - -- The cap region is `G_r = {λ_min(−A_M) > (4/(3k)) r M_3², rM_4 ≤ 3k/10}`. So depth failure is the layer - `0 < λ_1 ≤ D r U²` with `D = 4K²/(3k)` (P (7.1)): **the layer width is `∝ r/k`**. -- P (7.3) integrates the soft factor over that layer: `∫_0^{DrU²} λ_1(λ_1 + ErU) dλ_1 ≍ (Dr)³ ≍ (r/k)³` (times moments of - `U`). Combined with the `r²` from the other factors, the numerator `E_Q[W_r 1_{fail}]` is `≍ r² (r/k)³` — **if** the - axial factors are kept at their true size. -- P (6.1)–(6.2) bound `|α_M|, |α_S|` by `h/2 = M_3/2` (canceled-pivot). But P (5.2) says `α_M ≈ −6k`, `α_S ≈ 6k` when - `rM_4 ≪ k`. Bounding `6k` by `M_3/2` loses a factor `k²` in the numerator that the denominator does not lose: - P (5.4) gives `Z_r/r² → z_0 = (6k)² E[det(A_0)² 1{A_0<0}]`, i.e. **`Z_r ≍ k² r²`**. With the true axial size the `k²` - cancels between numerator and normalizer; with the canceled-pivot bound it does not. -- The fourth-derivative exception P (7.7) is `[10r/(3k)]⁴ E[(W/r²) M_4⁴] ≍ (r/k)⁴`, subdominant to `(r/k)³` for `r < k`. - -Heuristic conclusion (H): `1 − p_r(b,k,u) ≲ C min{1, (r/k)³} · (1 + |b| + k)^N` uniformly, for `r ≤ r_0`. -For `k` in a compact window this is the parent's `C r³`; the new content is the `k^{−3}` growth as `k ↓ 0` and the -saturation at `r ≍ k`. Physically: the whole cubic pair landscape `G_k` is `O(k)` in scaled units (every term of -`G_k` carries `k`), while third-order fluctuations are `O(1)`; when `k ≪ 1` the pairing survives only if the -transverse curvature beats the fluctuation, `λ_1 ≳ rM_3²/k`. - -## 4. Consequence if (H) holds: `ρ_rej(ℓ) = O(ℓ^{−1/4})`, and the strip splits at `−3/4` - -Insert (H) into §2 with the parent's majorant `H(b,k)` (P (13.4)) and `r³ = ℓ/k`, so `(r/k)³ = ℓ/k⁴`: - - ρ_rej^near(ℓ) ≲ ℓ^{−1/3} ∫_{ℓ/r_0³}^{∞} min{1, ℓ/k⁴} k^{−2/3} H̃(k) dk. - -The threshold is `k_* = ℓ^{1/4}`. Below it (`ℓ/r_0³ ≤ k ≤ ℓ^{1/4}`) the integrand is `k^{−2/3}` and the integral is -`≍ k_*^{1/3} = ℓ^{1/12}`; above it the integral is `ℓ ∫_{k_*}^∞ k^{−14/3} dk ≍ ℓ · k_*^{−11/3} = ℓ^{1/12}`. Hence - - ρ_rej^near(ℓ) = O(ℓ^{−1/3 + 1/12}) = O(ℓ^{−1/4}), - -and with the far part `Θ(1)`: **`ρ_rej(ℓ) = O(ℓ^{−1/4})`**. The rejected measure then has finite inverse moments -`∫ ℓ^q ρ_rej dℓ` for **`q > −3/4`**, narrowing the unresolved strip from `(−1, −2/3]` to `(−1, −3/4]` on the upper side. -(Checker `exponent_check.py` verifies the threshold algebra and both piecewise integrals exactly, and the contrast -scenario: a `k`-uniform `C r³` would give `ρ_rej^near = O(1)` and threshold `q > −1`.) - -## 5. Conjecture, and why the every-`d` lower-bound event does *not* settle it - -**Conjecture C7-τ.** `ρ_rej(ℓ) ≍ ℓ^{−1/4}` as `ℓ ↓ 0`, in every fixed `d ≥ 2`; the rejected measure's inverse-moment -threshold is exactly `q = −3/4`; in particular the unrestricted difference is **not** `O(1)`. - -The upper half is §4 given (H). The lower half needs a small-`k` failure mechanism of the right size. The direct -lower-bound events of the planar record and of PR #149 (soft eigenvalue pinned at `(3k/2) r`, third-order jets pinned -within `εk` of `O(k)` targets) have mass `∝ k⁴ r`, weight `∝ k⁴ r⁴`, normalizer `∝ k² r²`: they yield only -`1 − p_r ≥ c k⁶ r³` — a specific `O(k)`-geometry event that becomes rare as `k ↓ 0`. The conjectured dominant -mechanism is different: a soft eigenvalue at the **larger** scale `λ_1 ≍ rM_3²/k` with `O(1)` third derivatives -(mass `∝ r/k`, weight `∝ r⁴` with `k` cancelling in `6kr · (r/k)`, normalizer `∝ k²r²`, ledger `(r/k)³`). Proving -pairing fails with probability bounded below on that layer is the open obligation. - -## 6. Proof obligations (ordered; each is a bounded task) - -1. **k-explicit Theorem A (upper).** Re-run P §§5–7 keeping `|α_i| = 6k(1 + O(rM_4/k))` (not `M_3/2`) in (6.1)–(6.2), - `D = 4K²/(3k)` in (7.1)–(7.4), the `(r/k)⁴` term (7.7), and `Z_r ≥ z_* k² r²` from (5.4)–(5.5) with the `k²` - made explicit. Expected output: `1 − p_r ≤ C min{1, (r/k)³}(1 + |b| + k)^N e^{−c(b² + k²)}`-type bound on - `r ≤ r_0`, with the `b`-tail from P (13.2)–(13.3). This alone upgrades `o(ℓ^{−1/3})` to `O(ℓ^{−1/4})`. -2. **k-explicit lower bound.** A failure event on the layer `λ_1 ≍ rM_3²/k` with `Q^W`-mass `≥ c (r/k)³` for - `r ≤ ck`; the polygonal-path method should adapt with a model in which the third-order term dominates the - cubic, but the margins are no longer `O(k)`. -3. **Assembly.** DCT is not needed for the upper bound (it is an integral of a majorant); for the lower bound one - needs `A_r ≥ c > 0` on a positive-measure `k`-range near `k_*`, which P (10.3) gives on compacts only — a small - extension to `k ↓ 0` at fixed `b` is required (the `k²` in `z_0` is the only degeneracy). - -## 7. What this memo is not - -Not a theorem, not a review verdict, not a register change. Consistent with every accepted statement: the compact- -window `Θ(ℓ^{2/3})` loss (unchanged), Theorem U's far `Θ(1)` (unchanged), and `o(ℓ^{−1/3})` (implied by `O(ℓ^{−1/4})`). -It contradicts nothing in tree; it sharpens the open question's expected answer and names the exact steps. - -## Reproduce - - python -B -S exponent_check.py # prints RESULTS.json byte for byte; -O identical - python -B -S exponent_check.py --mutant M # exit 1 for M in {M1, M2} diff --git a/reviews/c7_total_k_scaling_recon_20260929/RESULTS.json b/reviews/c7_total_k_scaling_recon_20260929/RESULTS.json deleted file mode 100644 index e62f660fd..000000000 --- a/reviews/c7_total_k_scaling_recon_20260929/RESULTS.json +++ /dev/null @@ -1,23 +0,0 @@ -{ - "object": "CL-C7-TOTAL-K-SCALING-RECON-20260929-v1-CHECKS", - "passed": true, - "results": { - "compact_window": "O(l^(2/3)) recovered", - "ledgers": { - "conjectured_layer": "(r/k)^3", - "cubic_event": "k^6 r^3" - }, - "scenario_H": { - "moment_threshold": "q > -3/4", - "piece1": "l^(1/12)", - "piece2": "l^(1/12)", - "rho_rej_near": "l^(-1/3) * l^(1/12) = l^(-1/4)", - "threshold_k*": "l^(1/4)" - }, - "scenario_uniform": { - "moment_threshold": "q > -1", - "rho_rej_near": "O(1)" - } - }, - "scope": "exponent bookkeeping only; the heuristic (H) is not verified here" -} diff --git a/reviews/c7_total_k_scaling_recon_20260929/RETRACTION.md b/reviews/c7_total_k_scaling_recon_20260929/RETRACTION.md new file mode 100644 index 000000000..e96888c23 --- /dev/null +++ b/reviews/c7_total_k_scaling_recon_20260929/RETRACTION.md @@ -0,0 +1,14 @@ +# Retraction record — CL-C7-TOTAL-K-SCALING-RECON-20260929-v1 (withdrawn 2026-09-29, same day) + +**Scientific effect:** NONE. Author: Anthropic Claude (session `session_01NMeKEismAyeqgdB4sy2NJU`). + +The v1 memo (Math-#150, first commit `2fa23ba`) conjectured `ρ_rej(ℓ) ≍ ℓ^{−1/4}` for the unrestricted rejected +density and an inverse-moment split at `−3/4`. **That conjecture was wrong**, and the error was the author's: +the selection bound `min{1,(r/k)³}` was inserted against the parent's crude majorant `H(b,k)` (P (13.4)), which +does not carry the `k²` vanishing of the candidate intensity at small gap marks (`Z_r/r² → (6k)² E[…]`, P (5.4) — +the origin of the `k^{4/3}` in the (15.2) gamma integral). With that factor restored the near rejected density is +`O(ℓ^{1/4}) + O(1)`, and the unrestricted difference is `O(1)`, not divergent. + +The corrected result is the author-side candidate `frontiers/c7_total_bounded_20260929/PROOF.md` (Theorem K, +Corollary T), landed on the same PR. The v1 route is preserved there as the rejected checker mutant `M1`. The +retraction is recorded here rather than by history rewrite; the v1 bytes remain in the PR history. diff --git a/reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json b/reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json deleted file mode 100644 index 469b543ca..000000000 --- a/reviews/c7_total_k_scaling_recon_20260929/SOURCE_FILES.json +++ /dev/null @@ -1,30 +0,0 @@ -{ - "author": "Anthropic Claude", - "conjecture": "rho_rej(l) ~ l^(-1/4); rejected inverse-moment threshold q = -3/4; unrestricted difference not O(1)", - "delivered_under": "explicit owner instruction, claude.ai session 2026-09-29; OWNER_STOP left in place; no self-merge", - "files": [ - { - "bytes": 7234, - "path": "RECONNAISSANCE.md", - "sha256": "619fa94a4fa01489cb3f11a57fb5bce8bd23f697bd1b94fbe0dfee5c91e23452" - }, - { - "bytes": 3918, - "path": "exponent_check.py", - "sha256": "6245458fbc3f5866f7222bb0f398cbabd32fed128dbbae3c026ec577378c9d87" - }, - { - "bytes": 628, - "path": "RESULTS.json", - "sha256": "54e79903acfa3d721baffd2ad928aee5eee6bb805b0114238bdfb71608389ae9" - } - ], - "kind": "reconnaissance memo + exponent checker; conjecture with proof obligations; proves nothing", - "mutants": [ - "M1", - "M2" - ], - "object": "CL-C7-TOTAL-K-SCALING-RECON-20260929-v1", - "scientific_effect": "NONE", - "session": "https://claude.ai/code/session_01NMeKEismAyeqgdB4sy2NJU" -} diff --git a/reviews/c7_total_k_scaling_recon_20260929/exponent_check.py b/reviews/c7_total_k_scaling_recon_20260929/exponent_check.py deleted file mode 100644 index d100af937..000000000 --- a/reviews/c7_total_k_scaling_recon_20260929/exponent_check.py +++ /dev/null @@ -1,80 +0,0 @@ -#!/usr/bin/env python3 -# exponent_check.py — exact exponent bookkeeping for CL-C7-TOTAL-K-SCALING-RECON-20260929-v1. -# Python stdlib only; fail-closed; -O byte-identical. Mutants: --mutant M1|M2 -> exit 1. -# Verifies only the piecewise power-law integrals of RECONNAISSANCE.md §4 (symbolic -# exponent arithmetic in Fractions) — NOT the heuristic (H) and NOT any theorem. - -import json -import sys -from fractions import Fraction as F - - -def check(name, cond): - if not cond: - print("FAIL " + name) - raise SystemExit(1) - - -def power_integral_exponent(p, lo_exp, hi_exp): - """For int_{c1 l^lo_exp}^{c2 l^hi_exp} k^p dk with p != -1: the l-exponent of the - dominant endpoint contribution. Returns (exponent, which_end).""" - q = p + 1 - if q > 0: # dominated by the upper limit - return (q * hi_exp if hi_exp is not None else None), "upper" - else: # dominated by the lower limit - return q * lo_exp, "lower" - - -def main(argv): - mutant = None - if len(argv) == 3 and argv[1] == "--mutant": - mutant = argv[2] - elif len(argv) != 1: - raise SystemExit(2) - res = {} - - # Scenario (H): 1 - p_r <= C min{1, (r/k)^3}, r^3 = l/k -> (r/k)^3 = l / k^4 - # threshold k_* : l / k^4 = 1 -> k_* = l^{1/4} - thr = F(1, 4) if mutant != "M1" else F(1, 3) - check("T_threshold", F(1, 4) == thr) - # piece 1: k in [l/r0^3, k_*], integrand k^{-2/3}: int ~ k_*^{1/3} = l^{1/12} (upper end dominates) - e1, end1 = power_integral_exponent(F(-2, 3), 1, thr) - check("P1_exponent", e1 == F(1, 12) and end1 == "upper") - # piece 2: k in [k_*, inf), integrand l * k^{-4} * k^{-2/3} = l k^{-14/3}: int ~ l * k_*^{-11/3} = l^{1/12} - e2, end2 = power_integral_exponent(F(-14, 3), thr, None) - check("P2_exponent", F(1) + e2 == F(1, 12) and end2 == "lower") - total = F(-1, 3) + F(1, 12) - check("RHO_near_exponent_minus_1_4", total == F(-1, 4)) - # inverse-moment threshold for a density ~ l^{-1/4}: int_0 l^{q - 1/4} finite iff q > -3/4 - check("MOMENT_threshold_minus_3_4", F(-1, 4) + 1 == F(3, 4)) - res["scenario_H"] = {"threshold_k*": "l^(1/4)", "piece1": "l^(1/12)", "piece2": "l^(1/12)", - "rho_rej_near": "l^(-1/3) * l^(1/12) = l^(-1/4)", "moment_threshold": "q > -3/4"} - - # Contrast scenario: k-uniform C r^3 = C l / k -> integrand l * k^{-1} * k^{-2/3} = l k^{-5/3} - # on [l/r0^3, inf): int ~ l * (l/r0^3)^{-2/3} = l^{1/3} -> rho ~ l^{-1/3} * l^{1/3} = O(1) - e3, end3 = power_integral_exponent(F(-5, 3), 1, None) - check("CONTRAST_O1", F(1) + e3 == F(1, 3) and F(-1, 3) + F(1, 3) == 0 and end3 == "lower") - check("CONTRAST_moment_threshold_minus_1", F(0) + 1 == 1) # density O(1): finite iff q > -1 - res["scenario_uniform"] = {"rho_rej_near": "O(1)", "moment_threshold": "q > -1"} - - # ledgers of the two candidate lower-bound events (§5) - # planar/#149-type event: mass k^4 r, weight k^4 r^4, normalizer k^2 r^2 -> k^6 r^3 - check("LEDGER_cubic_event", (4 + 4 - 2, 1 + 4 - 2) == (6, 3)) - # conjectured dominant layer: mass r/k, weight r^4 (k cancels), normalizer k^2 r^2 -> r^3 k^{-3} - kexp = (-1) + 0 - 2 - check("LEDGER_conjectured_layer", (kexp, 1 + 4 - 2) == (-3, 3)) - if mutant == "M2": - check("LEDGER_conjectured_layer_mutant", kexp == -2) - res["ledgers"] = {"cubic_event": "k^6 r^3", "conjectured_layer": "(r/k)^3"} - - # compact-window consistency: k in [k-, k+] fixed -> min{1,(r/k)^3} = C r^3 = C l/k -> loss O(l^{2/3}) - check("COMPACT_consistency", F(-1, 3) + 1 == F(2, 3)) - res["compact_window"] = "O(l^(2/3)) recovered" - - print(json.dumps({"object": "CL-C7-TOTAL-K-SCALING-RECON-20260929-v1-CHECKS", - "scope": "exponent bookkeeping only; the heuristic (H) is not verified here", - "results": res, "passed": True}, indent=2, sort_keys=True)) - - -if __name__ == "__main__": - main(sys.argv) From 12eba3e99f54343d4d0dafe91b5b99844c682b48 Mon Sep 17 00:00:00 2001 From: Claude Date: Tue, 29 Sep 2026 20:48:18 +0000 Subject: [PATCH 3/3] C7-total v1.2: state the shifted-Gaussian eigenvalue integration with the frame fixed first Wording-only revision recommended in the OpenAI reviews 5357773376 and 5358141048: section 3 now says explicitly that, for fixed orthogonal O, ||A - m_A||_F^2 separates into shifted one-dimensional Gaussian factors, the lambda_2..lambda_m moments are polynomial in ||m_A|| uniformly over compact O, and the Haar integral is taken last; no orthogonal invariance of the shifted law is asserted. No theorem, constant or checker change; RESULTS.json bytes unchanged; SOURCE_FILES.json rebound to the new PROOF.md. Scientific effect: NONE. Co-Authored-By: Claude Fable 5.1 Claude-Session: https://claude.ai/code/session_01NMeKEismAyeqgdB4sy2NJU --- frontiers/c7_total_bounded_20260929/PROOF.md | 26 +++++++++++++++---- .../SOURCE_FILES.json | 4 +-- 2 files changed, 23 insertions(+), 7 deletions(-) diff --git a/frontiers/c7_total_bounded_20260929/PROOF.md b/frontiers/c7_total_bounded_20260929/PROOF.md index 1fe08e2a9..28748b086 100644 --- a/frontiers/c7_total_bounded_20260929/PROOF.md +++ b/frontiers/c7_total_bounded_20260929/PROOF.md @@ -1,6 +1,6 @@ # C7-total: a gap-mark-explicit failure intensity and the bounded unrestricted selection difference -**Object:** CL-C7-TOTAL-BOUNDED-20260929-v1.1 (supersedes the withdrawn v1 memo CL-C7-TOTAL-K-SCALING-RECON-20260929-v1; +**Object:** CL-C7-TOTAL-BOUNDED-20260929-v1.2 (v1.1 + the fixed-frame wording of the shifted-Gaussian step, see Revision history; v1.1 supersedes the withdrawn v1 memo CL-C7-TOTAL-K-SCALING-RECON-20260929-v1; see `RECONNAISSANCE.md` for the retraction). **Author:** Anthropic Claude (session `session_01NMeKEismAyeqgdB4sy2NJU`). **Disposition:** author-side proof candidate; **nonauthor analytic review required**. Scientific effect: NONE. @@ -69,9 +69,14 @@ Work on the typed support with `B = −A_M ≻ 0`, eigenvalues `0 < λ_1 ≤ … `U := J_r + ‖A_r‖_F + |b| + k ≥ 1`; then `h, M_4 ≤ K U` with `K` depending on `d, L` only. - **(3.5)/(7.2) with a shifted center.** The `Q`-density of `A_r` is Gaussian with covariance uniformly bounded above and below (P §3) and mean `m_A = E_Q A_r`, `‖m_A‖_F ≤ C(|b| + k)`: `density_Q(A) ≤ C_0 exp(−c_0 ‖A − m_A‖_F²)`. - Moments of polynomials in the eigenvalues against this shifted Gaussian are polynomial in `‖m_A‖`, hence in - `(1 + |b| + k)`. Below, the Gaussian factor in `λ_1` is dropped on the layer (as in P §7) and `λ_2, …, λ_m` are - integrated against the shifted Gaussian; this replaces P's centered `e^{−c Σ λ_j²}`. + The eigenvalue integration is performed **after fixing the orthogonal frame**: write `B = −A = O diag(λ) O^T` + and, for fixed `O`, `‖A − m_A‖_F² = ‖diag(λ) + O^T m_A O‖_F² = Σ_i (λ_i + q_i)² + (fixed off-diagonal squares)`, + `q = diag(O^T m_A O)`, `|q_i| ≤ ‖m_A‖`. The Gaussian factor in `λ_1` is dropped on the layer (as in P §7); the + remaining `λ_2, …, λ_m` integrals are one-dimensional *shifted* Gaussian polynomial moments, `∫ λ^p e^{−c_0(λ+q)²}dλ + ≤ C_p (1 + |q|)^p`, hence polynomial in `‖m_A‖ ≤ C(|b| + k)`, uniformly over the compact `O`; the Vandermonde has + already been replaced by the nonnegative polynomial `Λ^{m(m−1)/2}`, and the Haar integral over `O` is then taken. + This replaces P's centered `e^{−c Σ λ_j²}` without any `exp(C‖m_A‖²)` loss, and it does **not** assert that the + shifted matrix law is orthogonally invariant — only the centered covariance bound is. With these two replacements, P §7 runs verbatim. @@ -105,7 +110,7 @@ version above) and integrate `λ_1` first, as in P (7.4); every other factor is Now `a² ≤ 72k² + r²K²U²/2` gives `a² k^{−3} ≤ C k^{−3}(k² + r²) U²`, `a² k^{−2} ≤ C k^{−3}(k² + r²) U²` for `k ≤ 1`, and `a k^{−2} ≤ C (k + rU) k^{−2} ≤ C k^{−3}(k² + r²) U` since `r k^{−2} ≤ (k^{−1} + r² k^{−3})/2` (AM–GM). Moments -are polynomial in `(1 + |b| + k)` by the shifted-Gaussian remark and the `C⁴` moment of §2. So +are polynomial in `(1 + |b| + k)` by the fixed-frame shifted-Gaussian step above and the `C⁴` moment of §2. So E_Q[W_r 1_{depth}] ≤ C r⁵ k^{−3} (k² + r²) (1 + |b| + k)^N (k ≤ 1), @@ -172,6 +177,17 @@ is P §7 with `k` tracked. Everything else is the parent's own machinery. Not cl rate for `ν_cand` or `ν_eld` individually; any change to the parent's existential scope or to Theorem A's compact formulation; any statement about the *selected* measure's threshold (still `−2/3`); nonauthor acceptance. +## Revision history + +- v1 (withdrawn): `RECONNAISSANCE.md` retraction; the `ℓ^{−1/4}` conjecture is not revived. +- v1.1 (2026-09-29, blob `1fe08e2a…`): Theorem K / Corollary T with the saddle mixed term retained. +- v1.2 (2026-09-29): the shifted-Gaussian eigenvalue integration in §3 is now stated explicitly as performed after + fixing the orthogonal frame `O` (separable shifted one-dimensional moments, uniform over compact `O`, Haar integral + last), as recommended in the OpenAI reviews + [5357773376](https://github.com/d6g8k5htny-coder/Math-/pull/150#pullrequestreview-5357773376) and + [5358141048](https://github.com/d6g8k5htny-coder/Math-/pull/150#pullrequestreview-5358141048). Wording only; no + theorem, constant or checker change; `RESULTS.json` bytes unchanged. + ## 6. Review requested (non-Claude lane) (i) that P's canceled-pivot step (6.1) admits the `k`-dependent axial bound (5.2), and that the saddle bound keeps diff --git a/frontiers/c7_total_bounded_20260929/SOURCE_FILES.json b/frontiers/c7_total_bounded_20260929/SOURCE_FILES.json index feb6d5841..f100a3e3e 100644 --- a/frontiers/c7_total_bounded_20260929/SOURCE_FILES.json +++ b/frontiers/c7_total_bounded_20260929/SOURCE_FILES.json @@ -15,9 +15,9 @@ "delivered_under": "explicit owner instruction, claude.ai session 2026-09-29; OWNER_STOP left in place; no self-merge", "files": [ { - "bytes": 13464, + "bytes": 14982, "path": "PROOF.md", - "sha256": "e695c67a654a216ca4c72126319b2babdacbee0cda7e58fd2dfc0c67a6dcf8ae" + "sha256": "16c0efe372c1b1f26460aaa109287fed78e6fbd2dd00a0530998db296278dd01" }, { "bytes": 7736,