# Audit — the norm-as-price sweep: one load-bearing line (three pending remedies) and five loose prices, each with its exact value Jobless `audit` return of this run, 2026-09-17. Type `audit`, no job: `sahtool complete` needs an `attempt_id`, and this revises served documents rather than answering an assignment. ## 1. The class of defect, defined once Return #899 named the defect and #902/#905 settled its value: a **price** (the cost of an interface, or a norm that a consuming theorem reads) is quoted by an **inequality** whose slack is a factor of the object's incidental dimension — `sqrt(rank)`, a pair count, a `min(1,·)` — instead of by the **exact norm of the specific object**; the bound is then read as the value, and a verdict is taken from it. The test used here is therefore not "is a norm mentioned" but two questions: 1. is the quoted quantity an **inequality for** the object the theorem consumes, rather than the object? 2. does a recorded **verdict** depend on the difference? A line that fails (1) but not (2) is **loose**: it over-counts the cost, so it is conservative, and the audit's job is to record the exact value. A line that fails both is **faulty**. ## 2. What was swept The served documents were re-fetched for this audit, so every quotation below is against the bytes now served (sha256 of the fetched copy): | document | sha256 (first 12) | bytes | |---|---|---| | `research/OUTCOMES.md` | `78c5ea9f7f96` | 207 017 | | `research/RESEARCH-HANDOFF.md` | `b3f33f44c994` | 21 193 | | `research/QUESTIONS.md` | `07cadf7fb13f` | 601 467 | | `research/REFUTED.md` | `964e1cd6be99` | 571 | | `research/small-divisor-kernel.md` | `f27435bf29dd` | 20 892 | | `research/structural-literature-audit.md` | `22a1f41b0607` | 23 450 | | `research/grouped-divisor-moment.md` | `b0be809da298` | 20 601 | | `research/structured-dispersion-estimate.md` | `10da6db188a5` | 36 199 | | `research/IMPORT-MAP.md` | `035b44b90627` | 74 517 | | `research/THE-DIALS.md`, `TWIN-REDUCTION.md`, `ZONE-POSTULATE.md`, `moving-cutoff-parity.md` | — | — | Vocabulary searched, over every line of the above: `Frobenius`, `Hilbert–Schmidt`, `nuclear`, `operator norm`/`op-norm`, `sqrt(rank)`/`sqrt(M)`/`sqrt(D)`, `at the price of`/`priced as`/`price of`, `cost(s) is/of/x^`, `number of pairs`, `trivial bound`, `absolute mass`, `L1`/`L2`/`L^2` norm, mass. ## 3. Findings ### 3A. The one fault, and it is outside the documents The prescriptive **documents** now carry no verdict-flipping instance. The single faulty line in the whole prescriptive layer is in **return #759's R2 row** (echoed in route 49's register): > "…by Hilbert–Schmidt an arbitrary both-index coefficient is then admissible with the Frobenius norm > as its price, so R2 is one Cauchy–Schwarz away from the printed statements." The exact value of that price, now proved rather than measured (`report-audit-nucC.md`, verified in `work/r2-nuc-C-analytic.py`): ```text R(v) := ||A||_nuc / ||A||_F = 1 + C v + O(v^2), C = pi(1+rho) sqrt((U_2U_4-U_3^2)(V_2V_4-V_3^2)) / (U_2 V_2) = 0.1515 (1+rho) on the corpus's boxes. ``` So the interface's price is the **nuclear** norm, and it is `1 + 0.45 v` at `rho = 2` — an `x`-free constant — not `sqrt(rank)`. The `sqrt(rank)` reading was #899's own defect: it quoted an upper bound that is attained only for `v >> 1` and measured it in that regime (`X = 1e9` against `m ~ 4096`), which is not the corpus's (`v = Ax/(MN) <= 1`, `v = 1` at the box). **Three remedies exist and all three were still pending when this audit was written:** `research/RESEARCH-HANDOFF.md` (#899), `research/grouped-divisor-moment.md` §6.1 (#902), `research/structural-literature-audit.md` (#905). ### 3B. The five loose prices, with their exact values | # | document, line | quoted price | exact value | direction | |---|---|---|---|---| | 1 | `small-divisor-kernel.md` §3 item 2 (l.129–134) | "the only cost is the truncation height `(1+v)x^epsilon` and the L1 mass `f`" | `1 + C v + O(v^2)`, `C = 0.1515(1+rho)`; over-count `1/C = 6.6/(1+rho)` = **2.2×** at `rho=2`, **3.3×** at `rho=1/2` | over-counts; `v<=1`, so no power of `x` | | 2 | `small-divisor-kernel.md` verdict (l.9) | "Mellin separation … costs `x^epsilon` with L1 mass `min(1, A x/(M N))`" | same object as 1; exact mass `min(1, C v)` to first order | over-counts | | 3 | `OUTCOMES.md` l.1905 | same clause repeated for `Q-small-divisor-kernel` | same as 2 | over-counts | | 4 | `small-divisor-kernel.md` §4(b) (l.201–203) | "removing it costs, absolutely, at most (number of pairs)(coefficient mass)(M/x) ≪ `N^2 M/x = x^(14/25)`" | `<= 2 pi (A/(M N))(coefficient mass) = 2 pi v (coefficient mass)/x`; over-count `>= M/(2 pi v) = M/(2 pi)` | over-counts by a **power of `x`**, harmlessly | | 5 | `OUTCOMES.md` l.1904 | "removing it costs `O(x^(14/25))` absolutely" | same as 4 | same as 4 | **Row 4 deserves the exact reason**, because it is the only one where the over-count is a power of `x` and not a constant. The removed object is the reciprocity correction `e(theta R/(m c))` of (9), whose per-pair size is already bounded in the preceding sentence of the same paragraph; the price is then `2 pi` times that **per-pair** size times the coefficient mass, i.e. `2 pi v (mass)/x`. The recorded form instead multiplies in a **pair count** and a factor `M` — and `M` is a power of `x`. It is a valid `at most` (the line exists only to assert "far below every budget in §1", whose smallest budget is a block of size `x^(53/50)`), so the transfer of (5) and (12) is **strengthened**, not weakened; but a consumer who reuses `x^(14/25)` as the price of reciprocity is quoting something `>= M/(2 pi)` times too large. `research/grouped-divisor-moment.md` §2 carries the same `(1+v)`/`f` bookkeeping as row 1 and is the document rows 1/2 explicitly point to ("the same `(1+v)` and `f` bookkeeping as grouped-divisor-moment (6)"); the same exact constant applies there. `structured-dispersion-estimate.md` §5's "endpoint variation costs `f^2(1+A x/(M N))`" is the same object again as row 1. ### 3C. Checked and NOT an instance, with the reason * `OUTCOMES.md` (closed-route table): "a uniform fixed-power **operator-norm** saving for the complete dual-frequency matrix `S(t, lambda h; q)` … REFUTED at this full-spectrum scope … the Gram matrix is `q^2 I - q J` and **the norm is exactly `q`**". A norm is quoted, but *exactly*, and the row is a refutation. Correct as written. * `OUTCOMES.md` (closed-route table) and `coefficient-structure.md` §3, `Q-coefficient-structure`: the uniform fixed-power **L²-norm** saving is REFUTED by a derived lower bound (top-band energy `>= c D W log W`, nonsquarefree L² norm `<= sqrt(D) W^{1/4}`). A refutation, not a price. Correct. * `Q-left-divisor-signs`, `Q-signed-moment`, and OUTCOMES's Lemma I/II entries: the `sqrt(M)` quoted there is the **actual loss of the stated Cauchy inequality**, not a bound standing in for an exact norm. Correctly priced; the lane's question ("can the `sqrt(M)` be avoided") is a question about an inequality, not a mis-quotation. * `IMPORT-MAP.md` row 2's "tropical Perron–Frobenius" — a vocabulary collision, no norm is priced. * `SEARCH-CONVENTIONS.md`'s "operator norms" is a search keyword for the Pascadi route, not a price. ## 4. Scope and limits, stated plainly * This audit prices **costs**, not savings. Four of the five loose lines over-count a cost, which is the safe direction; a sixth kind of defect — a norm bound quoted as a **saving** — was searched for (the same vocabulary) and not found in the swept documents. * The fault of 3A is quoted from return #759, which this run read in an earlier window; the served **route 49** register repeats it. Neither is in the freshly fetched document set, so its *current* wording is quoted from the run's own record of #759, not from a fresh fetch of the return. * The exact constant `C = 0.1515(1+rho)` holds on the boxes measured (five shapes, `M` from 128 to 8192, `rho = 5/4 … 5/2`). Its dependence on `rho` is confirmed; the claim is not extrapolated to boxes outside that family. * Nothing here moves any exponent of the twin target. ## 5. What was deposited One revision, to the note that **owns** both prices and carries no pending revision: `research/small-divisor-kernel.md` (base `f27435bf29dd`) gains §4bis, which records (a) the closed form of the `(1+v)` price and (b) the exact per-pair removal price with its spurious pair count and factor `M`, and lists the same two objects where OUTCOMES and `grouped-divisor-moment` repeat them. The three documents that already carry the 3A correction (#899, #902, #905) are listed but not touched, since each has a pending revision on its own base. ## 6. Cheapest credible check for a reviewer ```bash python work/rev_sdk_price.py # base f27435bf, +53/-0, artifact written python work/r2-nuc-C-analytic.py # the exact constant: C_pred/(R-1)/v = 1.0000 on 8 boxes python work/fetch_docs.py . research/OUTCOMES.md research/small-divisor-kernel.md # confirms the two carriers' shas above ``` The reviewer should confirm, in this order: (i) `small-divisor-kernel.md` §3 item 2 and §4(b) quote the two prices as this report states, at the shas above; (ii) §4bis's block does not contradict either sentence's own conclusion; (iii) the constant `C_pred` reproduces `(R-1)/v` to four decimals.