# Job #1840 — the longest-run statistic bounds the correction's support, and the LOOSE support is not the problem Run `run_20260918_124244_BuHARA`, attempt `cb0975a6da2c042b30f596399df7e8c5`, session `7f304bb6c9081e7ebc657459`, department `dept_c326cb5ae203e5d0d94f8db1`, mode **general**, lane `g2-exponent`, type explore / purpose discovery, stage pursue, **route 67** (revision 2, active). Identity bound to this turn's chat: agent `base3-free-deepseek-flash`, model `deepseek/deepseek-v4-flash`, `X-Effort: unmeasured` (no effort/reasoning field is exposed by this app version; sources checked are recorded in `state/identity/run_20260918_124244_BuHARA.json`). Effort was **not** guessed. Rung: `verified` for the control reproductions and the structural statement, `measured` for the new sweep; per-claim wording below. ## 1. Answer to the route's question Route 67 asked: does #159's own largest untruncated `L` equal or stay below #161's `L(T_x,q)` at every shared `(x,q)`, and does the **LOOSE** form's support stay at or below `LMAX = 8`? * **Refined form:** yes, and the two sides are the *same* number at every pair this computer can reach. Measured on the true tile `T23` (`D = 7,952,175`, `G2 = 204`): the kill-graph longest component is `2` at `q = 29`, `3` at `q = 31`, `2` at `q = 37`. #159's own recorded field, read at source from return #159's `out/b.log`, is `max L = 3` at `T23 by 31` and `2` at `T23 by 37` — identical. #161's published diagonal reproduces **6/6** below, and its two higher entries (`L(T29,31) = 4`, quoted on the route) sit above my `L(T23,29) = 2` as they must. So `#159's largest L <= L(T_x,q) <= 4 <= LMAX = 8` wherever a number exists. * **LOOSE form (the actually open half):** its support is **not** near 8. The loose interior condition is per-gap, so a window of index `L` needs `g_{i+1}..g_{i+L-1}` all to qualify with `g mod q in {0, 2, q-2}`; the support is therefore `R + 1` where `R` is the longest run of *consecutive qualifying gaps* (the span condition only deletes starts, never adds them). Measured on `T23`: `R = 2` at `q = 29, 31, 37` (support `3`), and over **all 170 primes `29 <= q <= 1009`** the maximum is `R = 2`, i.e. support `<= 3`. Cross-checked against #159's own recorded `longest qualifying gap run` = `2` (`T23 by 31`) and `2` (`T23 by 37`) — identical; its other two folds record `3` (`T31 by 37`, `T37 by 41`), i.e. support `4`. **No measured or quoted pair reaches `LMAX = 8`.** * **Why, structurally (small proof, not an extrapolation):** for `q > G2(T_x) + 2` no gap of the tile can satisfy `g = q - 2` (`g <= G2`), `g = 0 (mod q)` (`0 < g < q`) or `g = 2` (impossible for a twin-admissible tile: two survivors differing by 2 would force a multiple of 3), so `R = 0` and the loose support is **exactly 1**. Verified for all 123 primes `q > G2(T23) + 2 = 206` up to 1009. For `q <= G2 + 2` the question is a finite measurement, and on the reachable tile it answers `R <= 2`. Consequence for the route's contribution: the producer's `LMAX = 8` truncation is **idle for both forms** at every pair with a number, and the reason the loose form did not break it is not that the loose condition is tight but that qualifying gaps are too sparse to run: `{0, 2, q-2} (mod q)` is a 3-class condition on a word whose gaps are multiples of 6. The route's registered failure condition (a pair whose largest `L > 8`, or a disagreement) did **not** fire. ## 2. What was computed (all exact integer, network-free, 30.87 s wall, one bounded `exec`, `exit_code 0`) Tile = twin-admissible residues mod `x#`, `D = T_x`; cyclic gap word `g`; `q` acts with `K = {0, q-2}`. Two statistics, both read off the same word by independent code: | statistic | definition | relation | |---|---|---| | `L_census` | longest component of the kill graph on nodes `(i,sigma)`, `sigma in K`, edge `(i,sigma)->(i+1,sigma')` iff `sigma' - sigma = g_i (mod q)` | `= #161's L(T_x,q) = #162's deepest census class = #159's max L` | | `R_loose` | longest cyclic run of consecutive gaps with `g mod q in {0, 2, q-2}` | `support(Q^loose) = R_loose + 1` | Controls (all reproduced, none tuned): | control | published | measured | |---|---|---| | `L(T_5,7), L(T_7,11), L(T_11,13), L(T_13,17), L(T_17,19), L(T_19,23)` | `2, 1, 2, 2, 2, 3` (#161 diagonal) | `2, 1, 2, 2, 2, 3` | | `T19` folded by `23`: runs of 3 | `62` (#162) | `62` (spectrum `{2: 11746, 3: 62}`) | | `D(T23)`, `G2(T23)` | `7,952,175`, `204` (#1632, #159 part 0.1) | exact | | `T23` by `31`, `T23` by `37`: `max L` | `3`, `2` (#159's own log) | `3`, `2` | | `T23` by `31`, `T23` by `37`: longest qualifying gap run | `2`, `2` (#159's own log) | `2`, `2` | New measurements: `R_loose` sweep over `q in [29, 1009]` (170 primes) — max `R = 2`, at `q = 29`; two independent implementations (pure Python list scan and a numpy index-difference scan) agree at `q = 29, 31, 37`; the large-`q` structural case holds for all 123 primes `> 206`. Reused department code: the tile constructor and the kill-graph census from `runs/run_20260917_115102_NivfOQ/work/job1634/src/job1634-checks.py` (reviewed), with its own two traps found and fixed here (see §4). ## 3. Index convention found (worth recording) `#161`'s diagonal is indexed by the **fold prime** and uses the tile at the prime *below* it: `L(T_{p-}, p) = 2, 1, 2, 2, 2, 3, 2, 4` at folds `7, 11, 13, 17, 19, 23, 29, 31` — the 7th entry is my measured `L(T23, 29) = 2` and the 8th is the route's quoted `L(T29, 31) = 4`. Reading the diagonal as `x = p` gives `3, 2, 2, 2, 2, 4` and fails 6/6 of the published values (first run of this script failed exactly that way). The diagonal is a statement about *consecutive* folds, not about `x = p`. ## 4. Own failures kept (so a successor does not repeat them) 1. **Doubled tile.** The reused `job1634` constructor starts `res = [1], P = 2` and loops over the primes *including* `2`, squaring the 2-adic modulus (`P = 4 * odd primorial`, two copies of the admissible set). Every **count** from it is exactly `2x`: its `T19 by 23` runs-of-3 came back `124` against the published `62`, `D(T23)` came back `15,904,350`. Maxima can also move. Fixed to the true modulus `x#`; the fixed build reproduces all controls. (First run: 14/22, all eight failures from this.) 2. **Isolated-node convention.** The reused census only enters a node with an outgoing edge, so an isolated node never appears in the spectrum; the statistic there is `1`, not undefined. The published second diagonal entry `L(T_7,11) = 1` is exactly such a case (its spectrum is empty). Without this, the run dies on `max()`. 3. **`Tprod` must skip `q = 2`** (`(2-2) = 0` zeroes the product) — a self-check on `D` reported `0` for every fold until fixed. ## 5. Prior-art update (this experiment) `web_search` **answered** this turn (control query `twin primes` returned the ordinary Wikipedia/MathWorld/arXiv set), so the channel is live and the negatives below are not channel failures. Queries: the control, plus `longest run of consecutive residues modulo p in a two-element set covering congruence Jacobsthal function maximal run`. Result: the external hits are about *consecutive quadratic residues*, *covering systems* and the classical one-class Jacobsthal/covering literature — none of them is a per-prime longest-run statistic on the fold word `T_x`, and none bounds the support of a correction term of this shape. This matches the route's own record (`#965`, `#968`): the coverage of this contribution is **internal** (`#159` refined form, `#161`/`#162` the run census, `#968` the definitional identity), and the exact remaining gap is the one measured here — the loose support at the folds `#159` actually proved something about, which this turn bounds as `R + 1 <= 4` by #159's own recorded values and `<= 3` by direct measurement on the reachable tile. Channels not queried this turn (bounded by the 0.5 h budget): arXiv API and OEIS. No external claim is made from this search. ## 6. Scope and limits * `T31 by 37` and `T37 by 41` are **quoted** from #159's log, not recomputed (their words are `D = 6,226,553,025` and `217,929,355,875`; this assignment's compute hint is `cpu_hours 0`, and the container's `alloc take` cap is 0 here). `T29`-indexed rows of #161 (`7 <= p <= 1009`) are likewise not rebuilt: the loose-support statement at `T29` is inferred only through the structural case `q > G2(T29)+2 = 260` (which is a proof, not a measurement) and through the quoted folds. * The `R + 1` identity is proved for the interior-run reading of `Q_L(theta)` quoted in the route itself; the threshold `theta` only removes starts (support is monotone non-increasing in `theta`), so `R + 1` is an upper bound for every `theta` and is attained for `theta <=` the smallest span at a longest qualifying run. * No claim about the exponent, the margin, `G2`, or any asymptotic. No novelty claim. The refined identity is #968's, checked numerically here, not re-proved. ## 7. Next step (distinct experiment) Measure the loose support where the route's proven numbers live: rebuild `T29` with the **constant-memory segmented sieve** already validated in this department (`2^24`-position numpy chunks give one `T29` period in ~11 s, `D(T29) = 214,708,725`), compute `R_loose` for `q in [29, 1009]`, and test `R + 1 <= 8` at the `T29` rows of #161's table that carry `L(T29,q) = 4`. Falsifier: any `q` with `R_loose >= 8` at `T29` (the loose support would then reach the producer's truncation, which the refined form does not). Budget 0.5 h, ~0.01 CPU-h. ## 8. Artifacts | file | sha256 (see `SHA256SUMS.txt`) | |---|---| | `job1840-checks.py` | the producer: tile, census, both loose-support implementations, 23 checks | | `job1840-checks.log` | `23/23 checks passed; all_pass=True`, `exit_code 0`, 30.87 s | | `job1840-checks.json` | machine-readable results incl. the 170-prime sweep | | `replies/ret159.json`, `replies/ret161.json`, `replies/route67.json` | the served records read at source | | `ops/*.json` | journals: register, the two failed `exec` runs (own failures §4.1/§4.2), the passing run |