# Job #4472 — explore, lane `dir-558`: `Q-varE-identification-0830` **Question.** Can either open step behind `lim Var/E = 0.45546` (the identification `delta*(X - X_dec) -> 0`, or the `theta = 2` mean-coefficient replacement) be proven — and if not, which single inequality does not close? **Answer taken forward, corrected.** Neither closes and they are one statement; the inequality that does not close is a *level-of-distribution* estimate for `y`-friable squarefree integers in arithmetic progressions, weighted by `lam1(e) = (1/e) prod_{p|e} p/(p-4)`, at bounded friability `u = log e / log y in (1.2, 2]`, **block-centered** (the printed prefix form is refuted) and needed only *on average in the modulus* against `lam0`. The registry row still names it in the older, stronger, refuted form (uniform pointwise `o(1)` for `d` up to `y^(4/5)`), so the row's answer is one refinement out of date. That is a documentation correction, not a mathematical one. ## What was done, in order 1. Read `research/README.md` (router), the `QUESTIONS.md` / `OUTCOMES.md` rows for this question, and the records they name: `research/history/staging/attack-0830-varE-identification.md`, `.../varE-theta2-step.md`, `.../varE-theta2-proof.md`, `.../recon-0830-smooth-aps.md`, `.../varE-limit-theorem.md`, `.../varE-exact-ladder-01.md`, and the two producers. 2. Searched online and read the nearest printed theorem **at the source**. 3. Rebuilt the served producer for this question with one added level, `x = 23` (its own record names that level as not recomputed). 4. Re-binned the dominant cell on the small part `d`, because the two printed bounds the record prices that cell against are *range*-restricted while the record's binned statement was only a *balance* statement. 5. Re-checked the load-bearing elementary claims in exact arithmetic (exhaustive where the record checked random samples). 6. Promoted the exact kernel into `private/lib/varE/` (17 unit tests, 4 negative controls). ## Findings, with the rung of each **F1 — new exact level `x = 23`, MEASURED.** `L = 23# = 223092870`, `y = 14929`, `ln y = 9.611`, 1745 primes in `[7, y]`, 4,561,931,320 descent nodes, 260.7 s wall on one node process (`node v22.20.0`), exit 0 with every corpus assert intact. Row sums: `MAIN = L` to 2e-6 relative; `REM = X = 121.485243`; and independently `delta*X = 0.364320` (the corpus reading of `varE-theta2-step.md` sec.4), `delta*X2 = 0.000236` and `delta*Xmix = -0.013876` (the 2026-09-05 rider's corrected values, asserted to 3e-6). So the level reproduces the record on all three published cross-checks. - `Xmix = -4.627014` against `-3.9404` at `x = 19`; `Xmix / ln y = -0.4814` against `-0.4900`. The record's declared falsifier — *growth at a new level* — does **not** fire; the coefficient falls. - The share of the mixed remainder at `n <= 2L` is `0.8081`, continuing the monotone fall `0.9861, 0.9160, 0.8437, 0.8257, 0.8178, 0.8081` over `x = 7, 11, 13, 17, 19, 23`. "82 % below `2L`" is a *falling* share, not a constant; the qualitative reading (the larger part is at `n <= 2L`) holds at every level. - The balanced pairs' remainder flips sign: `-0.0087` at `x = 19` to `+0.0018` at `x = 23`, so the unbalanced share reads `1.0005`. The magnitude is three orders below the cell, so "the obstruction is unbalanced" is only strengthened, but the record's exact `-0.0087` is an `x = 19` value, not a law. **F2 — new: the obstruction's mass sits at the *smallest* admissible moduli, MEASURED.** The dominant `mix2a` cell below `2L` was re-binned on `d = min(n0, e)` (the modulus `E_d` sums over, in the parent note's sec.4). Cumulative share of the cell's remainder: | `x` | `d < 10` | `d < 30` | `d < 100` | `d < L^(1/3)` | `d < (2L)^(1/3)` | cell REM | |---|---|---|---|---|---|---| | 19 | 0.3853 | 0.8487 | 0.9642 | 0.9856 | 0.9856 | -3.273753 | | 23 | 0.3960 | 0.8621 | 0.9746 | 0.9994 | 0.9994 | -3.824747 | `d = 7` alone carries `0.3960` of the cell at `x = 23`. Both totals reproduce PART C's `mix2a, n <= 2L` remainder independently (`-3.273753`, `-3.824747`), and the dlog cell total matches PART C to `< 1e-3`. This is the quantitative form of the record's own claim that the obstruction is *a short factor below every fixed power `L^delta`*: at `x = 23` the entire mass is at `d <= 100 = L^(0.223)`, and 86 % of it at `d <= 29`. **F3 — the nearest printed theorem reaches none of it, READ at the source.** Harper, *Simple Barban–Davenport–Halberstam type asymptotics for general sequences*, arXiv:2412.19644v1, Theorem 1 (read in the paper's HTML at the statement) requires `sqrt(2x) < Q <= x`. On this cell's dictionary (`Q = d`, `x = L/d`) that is `d > (2L)^(1/3)`: the theorem is silent on the `9856/9994` parts in ten thousand of the cell that F2 locates, and its covered sub-range is the top `13.1 %` (`19#`) / `13.7 %` (`23#`) of the log-modulus range `[1, L^(2/5)]`. The record's range objection is therefore not marginal; F2 and F3 together say the nearest theorem covers `< 0.06 %` of the measured obstruction at `x = 23`. **F4 — the fixed-modulus obstruction constant is right, VERIFIED.** `recon-0830-smooth-aps.md` sec.3.1a refutes the *printed prefix* hypothesis `(H_w)` with a modulus-7 character component. Re-derived here: the first block of `L(1, (./7))` is exactly `21/20`, all of the first 400,000 blocks are positive, `sum_{n>=11} 5/(n(n-4)) = 1207/2016` exactly (so `1 - that = 809/2016`), hence `C_chi >= (21/20)(4/5)(809/2016) = 809/2400 = 0.337083 > 1/3`. Evaluated: `C_chi = 0.90628` in its product form `prod_{p not 2,3,5}(1 + chi(p)/(p-4))` (`p <= 2e6`) and `0.90622` in the record's Euler-factor form; the bound is correct and 2.7x loose. `(1/6) C_chi^2 = 0.1369` is the limiting lower bound, positive. *Scope:* the finite evaluation confirms the constants and the block pairing; the diagonal-limit reading of the counterexample (the `y`-friable weight changes with `E`) is the record's and was not re-derived here. **F5 — the elementary identities hold exhaustively, VERIFIED.** `R_n(c) = [(r-c)^+ + (r+c-n)^+ - r^2/n]/L` agrees with the brute-force shift sum for **every** `(L, n, c)` with `L <= 25, n <= 40` (0 mismatches; the record checked 4000 random triples); `sum_{c mod n} R_n(c) = 0` exactly; `|R_n(c)| <= min(1, n/L)`; the `n > 2L` triangle form; `R_n(0) = r(n-r)/(nL)` (Fact A as a special case); and both forms of the mixed lag's CRT class, `c = 2 n0 n- a - 2 n0 n+ b` and `c = 2 n0 v` with `v^2 = inv(n0)^2 (mod n+ n-)`, on every coprime triple from `{7,11,13,17,19,23}`. **F6 — registry row.** `research/QUESTIONS.md` v10 `a3e07372` carries `Q-varE-identification-0830` at `PARTIAL` with the verdict of the 2026-08-30 note. That verdict predates `Q-recon-0830-smooth-aps` (the same TODO item 9): the printed prefix normalisation of the named inequality is refuted, the deduction needs only the `lam0`-weighted average over `d`, and accepted returns #921/#926 already control the zero/plus and zero/minus two-branch families at `Q <= X^(1/8)`, `y >= X^(1/3)`. The row is not *wrong* about which step is open; it names it in the form that no longer holds. Corrected row verdict text is in the return's `research.evidence_md`. ## The gap that remains, and what would falsify the reading The step still does not close, and nothing here bounds `G2`, `beta_2` or twin-prime infinitude. What is now sharper is *where* it is open: measured, the remainder is carried by `d = O(1)` — `7, 11, 13, ...`, not by moduli near `y^(4/5)`. Two consequences, one of each sign for the record: - the parent note's "short factor below every fixed power `L^delta`" is confirmed as the location of the mass (F2), and the uniform-in-`d`-up-to-`y^(4/5)` phrasing of the row is stronger than what the mass requires; - but a fixed-modulus statement is *not* automatically easier: at `u` bounded and `x = L/d -> infinity` with `y ~ sqrt(L)`, equidistribution of `y`-friable integers mod a fixed `d` is exactly the `u in (1,2]` register the recon note found no printed source for, so the gap is relocated, not closed. **Falsifiers, written before the next run.** (i) the small-`d` concentration is not a law: if at `x = 29` the share with `d <= 30` falls below `0.80`, F2's reading dies; (ii) if `Xmix / ln y` at `x = 29` exceeds `0.4900`, F1's no-growth reading dies; (iii) if any theorem is produced that gives the block-centered weighted mean-square estimate for `y`-friable `e` at fixed `d` and bounded `u`, F6's scoping is superseded. ## Files (all project-relative, scrubbed) - `outputs/job4472/attack0830_level23.js` — the served producer plus the `x = 23` level and the merged top bands (built by `make_level23.py` from the served copy, `code-sha256 85e05f41...`; only the LV table, two band-edge expressions and three guarded asserts differ). - `outputs/job4472/attack0830_dlog.js` — standalone log-binned `d`-split (`dlog_raw.txt`); `attack0830_dband.js` — the coarse `L^0.1 / L^(1/3) / L^(2/5)` binning. - `outputs/job4472/level23_raw.txt`, `dband_raw.txt`, `dlog_raw.txt` — raw stdout. - `outputs/job4472/window_kernel.py` + `test_window_kernel.py` — exact kernel and the `chi = (./7)` constant; promoted into the department's reusable Python library (`lib/varE/`, documented in that library's README; 17 tests, 4 negative controls). - `outputs/job4472/verify_job4472.py` — stdlib-only checker, 36 checks. - `outputs/job4472/check_chi.py` + `check_chi.json` — the F4 constants. - `outputs/job4472/REPORT.md` — this report. Rungs: F1, F2 MEASURED (exact finite computation); F3 READ AT THE SOURCE; F4, F5 VERIFIED (exact identities and finite evaluation); the reduction itself and the identification step remain OPEN; `lim Var/E = 0.45546` stays HEURISTIC.