Corpus audit of 1 October 2026: two-axis scores for all 175 notes, the uncorrected errors they expose, and where the drift sits

Markdown source · PDF

Result. Every maintained note was read in full and scored from 1 to 10 for interestingness and for correctness of its own derivations by Claude Opus 5.5 under one fixed rubric; Claude Fable 5.1 then reviewed the condensed statements of the 26 notes scored 6 or more for interest against their sources and the trap list of LLM.md §4. Three findings matter for the programme.

1. The spine is sound. The Newton recording results, the refinement and gauge insertion results and the strong-coupling constants score 8 or 9 for correctness; the reviewer found one false claim among the 26 selected statements (item (a) below) and no reintroduced trap.
2. Twenty-four notes carry uncorrected overclaims or inconsistent numbers (correctness 6 or below, four of them at 4). Twenty of the twenty-four belong to the Hamiltonian, flow and large-field mass-gap attack that preceded the 17 September change of direction, routes that the position note §3 lists as closed. The closure is recorded in the map; the route notes’ own texts keep the claims that the closure withdrew. That is the drift: supersession lives in the map and in LLM.md, while the superseded notes still read as live results.
3. Interest is concentrated. The scorer put 26 notes at 6 or above and none above 7; 70 of 175 notes are maps or calculations (kind counts below). The distributions are a calibration of what an outside expert would take from the corpus, under the rubric stated in §1, and settle nothing about novelty.

The two digests built from the scores, The most interesting results (26 notes) and Interesting and correct results (the 22 of them scored 8 or more for correctness), are published on the director’s personal site as phone-format PDFs with a landing page that tabulates every score: https://lxbifi11.bifi.unizar.es/navstokgap/. The repository keeps the scores here, in §4, and the corrections to make, in §3.

1. Method and reading labels

Six Opus readers worked in parallel on alphabetical batches of about 62,000 words each, each note read line by line, with no scripts for mathematics (project rule). The rubric’s anchors: interest 9 or 10 for a theorem with a surprising conclusion or a precise closure of a natural route, 7 or 8 for a solid nontrivial theorem, construction or decisive obstruction with explicit constants, 5 or 6 for useful supporting work, below that for bookkeeping; correctness 9 or 10 for complete proofs at the stated level with every hypothesis stated where used, 7 or 8 for essentially correct with minor gaps, 5 or 6 where a step is unjustified or claims outrun proofs, 4 or below for a substantive error or an unmarked superseded claim. Correctness was scored on the current text, so a note that was wrong and now carries a dated correction is scored on the corrected text. Each reader also wrote a 60 to 260 word condensed statement per note and a one-line reason for each score; the reasons are quoted in §3 for the low scores.

The reviewer (Fable) read the 26 selected statements against the notes’ lead sections and theorem statements, checked the arithmetic of the constants that appear in them, and checked them against the 26 traps of LLM.md §4. Items marked verified below were checked by the reviewer from the note’s own formulas; items marked reported are the scorer’s reasons, quoted and unverified here. Reading label for this note: scorer full-read of every note; reviewer passage-level reads of the 26 selected notes and grep-level checks of the notes in §3.

Scores are a model’s reading with reasons on file. They are not refereeing, and a high correctness score means a complete derivation at the level the note states, which may be conditional, formal or a reading.

2. Distributions

Interest: 1: 3, 2: 8, 3: 41, 4: 58, 5: 39, 6: 21, 7: 5. Correctness: 4: 4, 5: 6, 6: 14, 7: 27, 8: 59, 9: 65. Kind: calculation: 44, construction: 13, exploratory: 5, map: 26, obstruction: 24, reading: 12, theorem: 51. Status as read by the scorer: conditional: 18, conjectural: 4, mixed: 39, none: 16, proved: 96, superseded: 2.

Selected for the digests (interest 6 or more): mark-cost-and-statistical-floor, planck-gap-paper and planck-gap-probabilistic at 7; the other 23 at 6. Of the 26, the four below 8 for correctness are planck-gap-paper (7), four-dimensional-composition (7), su2-midplane-small-field (7) and planck-gap-derivation (6).

3. Errors and overclaims found

3a. Verified by the reviewer

1. planck-gap-derivation, boxed area forms off by a factor 2. The note’s theorem \(F^2\tau^3>9m\kappa\) with \(\tau\Delta E=F^2\tau^3/(2m)=(3F/v)A\) and \(A=vF\tau^3/(6m)\) gives \(A>3v\kappa/(2F)\); the box states \(A>3v\kappa/F\). Corollary 3’s two-mark upper value is likewise \(6v\kappa/F\), stated as \(12v\kappa/F\). The theorems themselves (constants 9 and 36) hold. The digest on the site carries the corrected values and a reviewer’s note. The interval model is in any case superseded by the Gaussian mark model of mark-cost-and-statistical-floor and the Planck paper, where \(\kappa\ge\hbar/2\) holds for every probe state, and the note should open by saying so.

2. strong-coupling-target-box, the \(q=4\) tolerance column disagrees with the note’s own criterion. From \(e\,\eta\,2^q\le0.15g^2-384e/g^2\) (the note’s own criterion) one gets \(\eta\le0.105\) at \(g^2=100\) and \(\eta\le0.570\) at \(g^2=200\); the table prints \(0.29\) and \(0.62\). The \(q=8\) column checks.

3. small-field-step-decay-and-threshold, abstract and body disagree on constants. The abstract quotes an error density \(10^{-12}\); §3 derives \(1.1\times10^4\) and \(10^{-13}\). The scorer also reports that §3 sums an \(\ell^\infty\) decay with the \(\ell^1\) product formula, undercounting by about 24; that step is reported, unverified here.

3b. Reported by the scorer, with the scorer’s reason

• flow-instability-large-field (I 5, C 4): Proposition 1 (symmetry of \(M\)) is correct. The sharpness claim fails on the note’s own formula: with the Landau level \(gB\) included, \(\omega^2=k_\parallel^2-gB\), so the growth rate is \(gB\), half the Duhamel exponent \(2\|G\|\). A symmetric operator of norm \(\|M\|\) need not attain \(+\|M\|\). Section 4’s ‘sharp’ crossover inherits this; the text carries no correction.
• large-field-action-lower-bound (I 3, C 4): Proposition 1 is correct in the continuum. Proposition 2 fails: a field constant on the block and zero outside violates the Bianchi identity, and flowing at radius equal to the block size changes the block integral by order one, not \(O(a/\ell)\). Section 3’s descent argument is a non sequitur. Section 4 imports the unsupported ‘sharp \(e^{2\eta}\)’ claim.
• mass-gap-conditional-theorem (I 4, C 4): The decay-to-gap transfer and rate bookkeeping in §3 are fine. But §4’s ‘nothing else is missing’ and ‘a proof consists of H1 and H2’ overclaim: continuum existence, OS axioms and nontriviality are not implied. Existence of the renormalized interaction and gauge-theory strong mixing are glossed. The later rejection of physical-coupling H2 certification is not marked.
• schur-error-ultraviolet (I 5, C 4): Lemma 1’ and the oscillator derivative check. Section 4’s own sum gives theta ~ g^2 (Lambda L)^3, an extensive vacuum-shift slope, not the stated g^2 (Lambda L)^2, so the regime g << 1/N_s in Section 5 appears wrong (g << N_s^{-3/2}). The Berry coefficient seems to drop the four charge/polarization copies. The fixed-lattice theorem is asserted but unproved.
• gapped-set-critical-coupling (I 4, C 5): Proposition 1 is fine. Proposition 2’s proof is heuristic: it equates finite-volume gap closing with infinite-volume degeneracy or diverging length, and ignores massless phases that fill intervals of coupling. Section 4 calls the compact \(U(1)\) transition second order with a Maxwell limit at \(g_c\), against lattice evidence of weak first order. The dated Wilson \(SU(N\ge5)\) precision is present.
• ground-state-measure-transfer (I 4, C 5): Proposition 1 is standard and correct. The claim that isotropic Wilson-measure estimates hold ‘verbatim’ ignores that the Hamiltonian limit is the anisotropic \(a_t\to0\) measure. The bound \(\mathbb P\lesssim e^{-c\eta^2/g(\ell)^2}\) in section 3 is a heuristic presented as available. The section 4 competition is marked superseded in part; the action display carries a stray factor.
• large-field-entropy-count (I 4, C 5): The Landau degeneracy, the longitudinal disc count, \(\mathcal N=\eta^2/(8\pi^2)\) and the ratio \(2\pi^2/g^2\) are arithmetically correct. The convergence claims in §§3–4 (‘settles’, sum converges for \(g^2<2\pi^2\)) rest on an unproved bounded-factor-per-mode premise and on counting only constant-background modes. Proposition 1 is trivial.
• mass-gap-position (I 4, C 5): Internal numbers disagree: the strong threshold appears as g^2>=176, >444 and >=1056; the weak side as 1/g^2>~102 and as g^2~10-12. ‘Large-field region closed in measure’ exceeds the cost and entropy ingredients cited. ‘Exhaust the methods built from operator norms’ generalizes from five specific closures.
• strong-coupling-target-box (I 5, C 5): The criterion e eta 2^q <= 0.15 g^2 - 384e/g^2 and the q=8 column check. The q=4 column does not: at g^2=100 the criterion gives 0.105, the table 0.29; at 200 it gives 0.57, the table 0.62. Tolerances ‘divided by 24 at the listed couplings’ fail, since 24*384e/g^2 exceeds 0.15g^2 there. The Proposition claims more than the cyclic-subspace criterion.
• typical-field-strength-window (I 4, C 5): The free-field integral, the range condition \(\kappa^2\ge g/4\pi+1/2\) and the Gaussian tail estimate check, with \(\hbar\) absorbed into \(g^2\). The claims that the Hamiltonian route terminates and that no measure exists ignore the vacuum state’s configuration distribution. A top box flags the lattice correction, but the title, section 4 and section 5 still state the superseded conclusion.
• abelian-misses-the-box (I 4, C 6): b_0=11/(16 pi^2) and 2 b_0 log 2=0.0966 check. The U(1) rerun of the expansion is asserted ‘verbatim’ without the threshold computed. ‘Gapped set is (g_c,infinity)’ (intro, §2) overclaims: no monotonicity or single transition is proved. ‘All three isolations are b_0 statements’ (§3) is an unlabelled reading.
• action-floor-yang-mills-gap (I 5, C 6): Prop. 1, Thm 2, Prop. 4 and the Bohr-Sommerfeld levels check. Remark 3 counts only the x-valley (both valleys double the constant; growth law unaffected). Prop. 5’s d=4 row drops 1/g^2, which by the table’s own convention has action dimension, so ‘no constant at all’ in d=4 is inconsistent.
• agmon-ground-state-suppression (I 3, C 6): Lemma 1 and Corollary 2 are correct; Corollary 2 with rho=(1-delta)d needs epsilon tied to delta, left implicit. The distance step (§2) is a labelled scaling argument. The correction box flags that only the global excess is controlled, yet §3-§4 still state local consequences (finite exponential local moments) that do not follow.
• confinement-scale-bands (I 5, C 6): Table entries (r_0/a, m a=4.21/(r_0/a), sigma a^{2=1.35/(r_0/a)}2), g^2=6/beta, the 444 threshold and the area-law values check; a dated caveat flags the softened glueball near beta=5.7. Treating a coarse Wilson loop as polymer activity (§2b) is heuristic. ‘Six blocking steps is the whole problem’ overstates: the weak side’s ‘perturbative in kind’ doublings are unproved.
• flow-before-decimation (I 4, C 6): The invariance proposition is correct as stated. The sweeping claim that every norm-based criterion is exhausted outruns it, since a conjugated Hamiltonian can be split differently. The addendum’s local \(e^{-\lambda n}\) ground-state suppression is superseded by the later Agmon-global correction and is not marked as such here.
• flow-jacobian-truncation-error (I 5, C 6): Proposition 2’s Kato-Duhamel bound is correct. Proposition 1 handles the gauge-fixing term loosely (‘to the same order’). The truncation error is a kernel-tail heuristic, not an operator relative bound. The Sharpness box repeats the instability note’s unsupported claim of growth rate exactly \(2\|G\|\). The lattice-step pessimism is correctly flagged as corrected.
• intermediate-region-finite-verification (I 5, C 6): The decay-to-gap step via reflection positivity is sound. The lead’s claim that the problem ‘reduces to’ two numbers is weakened only in §4b. The lead says gauge invariance is no obstruction, while §5 admits Elitzur empties the single-link form; the tension is left unresolved. Box sizes are corrected in a later note, flagged here. Literature is at metadata level.
• large-field-operator-inequality (I 5, C 6): Propositions 1 and 2 (min-max and the ground-state-transform trial state) are correct, and the IMS arithmetic giving \(C_L=32\pi^2c_0^2/3\) checks. The link-derivative bound behind \(|\nabla f_B|^2\) is heuristic. The claim that every Hamiltonian route loses a surface term is shown for one crude trial construction only; window constants are crude.
• lattice-truncation-uniform (I 5, C 6): Coupling cancellation in the lattice flow and the compactness bound are correct, and exponential locality of the linearized flow over unit time is standard. The Proposition’s proof is a sketch: a kernel tail is equated with a relative operator error, the Jacobian potential from \(\rho_t\) is not treated, and ‘K=3 suffices’ is asserted without values for \(C,c\).
• planck-gap-derivation (I 6, C 6): Lemmas 1-3, Theorem 1 (constant 9) and Theorem 2 (36) check. The area forms are off by a factor 2: since (3F/v)A = tau Delta E, the bound is A > 3v kappa/(2F), and Corollary 3’s upper value is 6v kappa/F. The note omits the later reinstatement of a quantum kappa = hbar/2.
• small-field-step-decay-and-threshold (I 5, C 6): The soft-constraint gap 2/9 at a>=130 and the rate 1.7e-3 check, as do the strip facts (a),(b) of §2 for a one-direction shift. §3 sums an l-infinity decay with the l1 product coth^4, undercounting by about 24. The abstract quotes 2e4 and 1e-12, while §3 and its table give 1.1e4 and 1e-13. ‘Closed’ outruns a heuristic estimate.
• strong-coupling-threshold-explicit (I 5, C 6): The Kotecky–Preiss chain checks: degree 80, c~220, rho<=8e-4, t_0>=7.1, e^{64 t_0}, beta_*~e^{-465}, g_0^2~1e101. §3’s 40–70 is an estimate that multiplies the per-site beta=54/g^4, which already sums three plaquettes, by a further coordination 12. The text does not point to the later rigorous 388/79. The weak-side figure 1/g^2~1e2 predates the small-field note’s 1e-12.
• su3-constants (I 3, C 6): b_0=11/(16 pi^2), 51/121, 54/g^4, the 8/3 g^2 flux loop, v<=96e c/g^2, Haar variance 1/2, the Delta V identity and the doubling table check. The blocking paragraph writes beta_direct=24/g^4; the cited note’s 32N/(C_2 g^4) gives 72/g^4, which is needed for 3M^2/8. ‘Equivalence of T2’ with no zero-temperature transition’ is asserted as standing without proof.
• what-would-unblock (I 4, C 6): Most citations and characterizations are accurate at metadata level. The infrared-bound paragraph gives the wrong reason: Froehlich–Simon–Spencer cover O(N) spins, so failure stems from the non-quadratic plaquette interaction. Section 6 states as fact that coarse effective interactions enter the completely analytical set. Item 1 is half-retracted in place while the opening still says ‘exactly two things’.

3c. Smaller slips reported in notes scoring 7 or above

The scorers’ closing reports name slips that leave each note’s main result standing. All are reported, unverified here, except where marked.

• three-dimensional-gap-one-function and low-dimensional-mass-gap (§7, the Göpfert–Mack quotation): the Debye-mass formula \(m_D^2=(2\beta/a^3)e^{-\beta v(0)/2}\) with dimensionless \(\beta\) carries the wrong power of \(a\); \(a^{-2}\) is right (verified here by dimensions: with dimensionless \(\beta\) the right-hand side has dimension length\(^{-3}\) and a Debye mass squared needs length\(^{-2}\)).
• newton-mark-floor: the three-mark side remark should give \(F^2\tau^3>64\,m\Lambda p\), stated as \(128\).
• planck-gap-probabilistic: Proposition 4 uses a Helstrom constant \(x^2/4\) where the bound is \(x^2/2\) (constant only); the aperture hypothesis should be stated on the hybrid states.
• rotation-composition-universality: the middle expression of eq. (8) has \(+i\) for \(-i\); the final expression and conclusion stand.
• weak-coupling-feshbach-reduction: in the order table the \([a,a]\cdot[\tilde A,\tilde A]\) part of \(W_3\) is quadratic in the fluctuations and of order \(g^{4/3}\), stated as \(g^{7/3}\).
• wilson-strong-coupling-explicit: the all-representation weight bound needs \(|\int\prod\chi_r|\le\prod d_r\), giving \(\sum d_r^2c_r=e^{\beta_W}\); the threshold 1056 stands.
• finiteness-half-flowed-susceptibility: an inline correction removes the \(g^2\) prefactor, but the summary and status table were not updated.
• dimension-ladder: one sentence states the Layer 1 conjecture as settled.
• lieb-robinson-kogut-susskind: the §4 condition should read \(g^{8/3}<N\).
• holography-lowest-dimensions: the entanglement asymptotics should carry \(+1\) in place of \(\log(1/d_1)\).
• sun-midpoint-centre §6: a broken \frac renders as “rac13”; su2-midplane-small-field: §9 is unrefereed and §§27–31 are internally checked only (status, recorded in the note).

Two items the scorers could not check: the Newton Project paragraphs cited by newton-NATP00385-audit, and the proof behind the routing stub ordered-beam-preparation, which sits in papers/collision-action-relaxation.tex.

3d. Where the drift sits

Of the 24 notes above, 20 are mass-gap route notes written before 17 September (Hamiltonian flow, large-field region, Agmon, blocking, truncation, Schur and threshold notes); mass-gap-position §3 closes their routes and LLM.md §4 records the traps, but the notes themselves still state their conclusions in the lead and keep inconsistent thresholds (176, 444 and 1056 for the Wilson strong-coupling gap appear with different labels across mass-gap-position, su3-constants and strong-coupling-threshold-explicit). The other four are planck-gap-derivation (item (a)), action-floor-yang-mills-gap (a \(d=4\) row that drops \(1/g^2\)), abelian-misses-the-box (a gapped set stated without monotonicity) and what-would-unblock (a wrong reason for the failure of infrared bounds). The repository’s own convention for a withdrawn claim is a dated correction box at the top of the note; applying it to the 24 notes, with the unified thresholds, would remove the drift without deleting the record.

4. Scores for every note

Interest (I) and correctness (C) from 1 to 10; status and kind as read by the scorer.

5. Consequence for STATE

The two proof goals are unaffected: the Newton recording spine and the refinement and gauge spine are the notes the scorer could follow to the end. The pending work this audit creates is corrective: dated correction boxes in the 24 notes of §3b, the slips of §3c, the factor-2 fix in planck-gap-derivation, the \(q=4\) column in strong-coupling-target-box, the abstract of small-field-step-decay-and-threshold, and one set of strong-coupling thresholds with their labels in mass-gap-position and su3-constants. STATE gains one line pointing here.