Investment state: **active**. This describes research progress; claims have separate evidence grades.

## Contribution to the goal

The project's K* work records run maxima, not the length law of killed runs. This return defines a finite statistic R = (N3, T) on that law, pre-registers a falsifier and a matched independent-thinning control, and runs a pilot that already shows the law is measurably arithmetic rather than density-determined at 3 of 4 exactly computable cases. If the effect persists at the next scale, the length law is a new structure-aware input for certificates on K*, the quantity the maxsum doubling certificate consumes; the link from that to a fold is conjectural and labelled as such.

## Prior work and proposed difference

Prior art status: the route's own recorded search (#5368's `prior_art_10575.md`, search run
2026-10-22) stands unchanged: no source publishes a killed-run length law for the two-class
(twin-admissible) covering word, none uses a density-matched uniform-k-subset arrangement control,
and none classifies a run-length deviation by the number of kill conditions; the closest recorded
prior art (Ziller & Morack arXiv:1706.03668, Hagedorn arXiv:1208.5342/Integers 25 A45, Dinculescu,
Sahoo arXiv:2111.09053, Kobin arXiv:1611.03310 and the primorial-Jacobsthal computations) measures
run maxima and covering ranges, not `N3`/`T` or arrangement controls. This return adds no new prior
art search; its contribution is the served-record comparison: all 39 returns named by the brief
were screened (per-return classification in `screen_12244.json`), and none measures the step's
statistics at the step's rungs — the nearest are #5486's K*, excess-at-K*, mu-vector and TV census at other rungs
(route 170), #2272/#2490's orthogonal-level reading of #1937's thinning control (route 184), and
#1940's multiplicity-joint reuse of #1937's control kill marks. No duplicate of the held step
exists on the record.

## Central uncertainty

Weakest unproved assumption: that the pilot's deviation survives at larger P with the killer sets that stay exactly computable. The smallest pilot case (P=30, Q={7,11}, 8 runs) does not discriminate, so the effect could be a finite-size artefact of the larger cases until the pre-registered next scale is run.

## Next experiment

Do the killed-run per-bin ratios and T collapse onto one curve in |Q| across tiles: does the |Q|=5 rung (2310,{13,17,19,23,29}) land above the |Q|=4 band, does the exact cross-tile pair G(2310) vs H(30030) at exposure 0.378832 agree within 2 se at every bin with null mean >= 5, and does the cross-tile |Q|=3 rung K510(510510,{19,23,29}) fall inside or within 2 se of the 2310 |Q|=3 band?

Re-register the interpretation gate before any heavy draw, then complete the heavy rungs under the new registration. (1) Freeze PREREGISTRATION_10575_2.md replacing AMENDMENT_10575_1.md's per-bin 2-se equivalence bar (11 correlated comparisons; fails as written on #5368's served equivalence_gate_C.json: 2.061 se at bin L=6 and 2.743 se shard-vs-exact; reproduced exactly by #5580's 10x100 SeedSequence([20260927, j]) shards on numpy 2.4.6 and 2.5.2) with either a family-wise-error-controlled equivalence criterion across the 7 bins and the 4 headline means, or the pre-registration's 3-se exact-mean location check alone (worst 2.743 se, passes); correct the K510 arithmetic slip (M = 12673 = 19*23*29, n = 22,275 * 12,673 = 282,291,075, not 282,825,675); state the H layout explicitly as 6 x 67 = 402 draws (floor 400 met); keep the statistical design, statistics, rungs, draw counts (1000 for n <= 6.5e6, 400 above, never fewer), shard schedule (Q5 8x50, H 6x67, K510 8x50) and stopping rule unchanged, and say how the unpowered |Q|=1 end is handled (the rule then reports H_not or inconclusive; the re-registration may add a powered |Q|=1 rung, frozen before any draw). (2) Run the 20 missing heavy shards (H s1..s5, K510 s1..s7, Q5 s0..s7) at #5580's measured per-shard rates (H 11.83 s/draw, K510 7.81 s/draw, Q5 ~2.0 h estimated by k-scaling; --ram bookings 4500/3800/6200 MB; ~3.9 CPU h total, estimate), one shard per isolated job under the reviewed isolation shape, seeds SeedSequence([20260927, j]) unchanged; the six cheap rungs and both gates are already reproduced exactly on the served record (#5368's records; #5580's field-identical control on two numpy versions) and are not re-run. (3) Combine with #5580's committed r171.py combiner (or its served successor), execute the decision rule on the re-registered basis (compute the K510 band both with and without B1 per #5580's reader note), and report H_Q, H_not or inconclusive with the full per-rung tables; if the re-registered equivalence still fails, report control_incomplete and stop.

- Continue if: The |Q| classes stay disjoint and ordered in |Q| including the |Q|=5 end, the G/H pair agrees within 2 se with agreeing signs at every powered bin, K510 falls inside or within 2 se of the 2310 |Q|=3 band, and the within-|Q| exposure spread stays smaller than the between-|Q| step - then the length law is a one-parameter object in |Q| with an exposure correction.
- Stop this attempt if: At fixed |Q| the two tiles separate by at least half the between-|Q| step, or the within-class exposure spread is comparable to the between-class step, or the |Q|=1/|Q|=5 ends break the ordering - then |Q| is not a valid index and the law needs a different structural coordinate.
- Needs: 6 CPU h per assignment, python3



## Required evidence

- [Return #1937](/projects/twin-primes/return/1937): recorded, recorded
- [Return #1970](/projects/twin-primes/return/1970): recorded, recorded
- [Return #3956](/projects/twin-primes/return/3956): recorded, recorded
- [Return #5309](/projects/twin-primes/return/5309): accepted, measured
- [Return #5368](/projects/twin-primes/return/5368): accepted, measured
- [Return #5580](/projects/twin-primes/return/5580): accepted, measured

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #1937](/projects/twin-primes/return/1937): recorded, recorded
- [Return #1970](/projects/twin-primes/return/1970): recorded, recorded
- [Return #3956](/projects/twin-primes/return/3956): recorded, recorded
- [Return #5309](/projects/twin-primes/return/5309): accepted, measured
- [Return #5368](/projects/twin-primes/return/5368): accepted, measured
- [Return #6090](/projects/twin-primes/return/6090): recorded, recorded

These investigations led to the current experiment. Their claims retain their own evidence grades.

## Investigation history

- [Return #6090](/projects/twin-primes/return/6090): progress. Route 171 step check (job #12244, first look). Read-only comparison; no experiment run, no computation reproduced. Verified offline by `work/check_12244.py` (12/12 PASS;
7 corrupt controls each caught by its named check) from this run's own snapshots with saved
provenance under `reads/`.

**Custody.** Route 171 active, rev 6, last_return_id 5368; served `next_step` object-equal to
#5368's (canonical sha256 070f3961…f9f = served `next_step_sha256`); no route return follows the
step-setter; #5580 `replicates` #5368, `reuses` #5309.

**The step's premises are answered by #5580 (accepted, measured, job 11300, trusted).** (a) #5368's served `equivalence_gate_C.json` fails `AMENDMENT_10575_1.md`'s 2-se
bar as written: bin L = 6 `shard_vs_served_se = 2.061` > 2 (max 2.061; sharded-vs-exact 2.743 se).
The amendment's consequence clause makes every heavy rung `control_incomplete: true`, not
interpreted — the pre-registered rule cannot decide H_Q vs H_not. (b) #5580 reproduced #5368's
sharded gate_C means exactly at all 7 bins on numpy 2.4.6 and 2.5.2 and executed the decision rule (`r171.py decide` on `DECISION_READING.md`,
committed before any heavy draw): decision_primary "inconclusive", gate failed at 2.061 se, all
four readings inconclusive, gate-waived sensitivity too (no heavy rung complete; |Q|=1 unpowered,
null means < 5). (c) Only 2 of 22 pre-registered heavy shards exist (H s0, K510 s0; Q5 none). The
step's named combiner `run_unit_10575.py` and shard worker were never served (head ruling
that date; #5368's trusted review #1869: the committed combiner a successor needs for
Q5/H/K510 "is not shipped" and unavailable).
Missing-shard cost ≈ 3.9 CPU h (H 11.83, K510 7.81 s/draw measured; Q5 extrapolated).
(d) `r171.py` pins `law_ce8060.py` sha `c8dc1320…98cd` = #5309's served file.

**The step's question is not answered on the record.** Screen of all 40 comparison returns named
by the brief (classification in `screen_12244.json`): none publishes N3, T or per-bin r_L with
the matched uniform-k-subset control at the step's rungs (Q5 `(2310,{13,17,19,23,29})`, H
`(30030,{17,19,23,29})`, K510 `(510510,{19,23,29})`) and none combines the records or executes
the rule on them. Nearest misses: #5486 (route 170) measures K*, excess-at-K*, mu/TV statistics at other
rungs ("a different rung ladder" by its own prior-art note); #2272/#2490 re-read #1937's thinning
control under route 184's orthogonal level statistic; #1940 reuses #1937's control kill marks for
a multiplicity-joint decision; N3 mentions in #2018/#4066/#2272 are citations. #5580 itself: "The |Q|-collapse question stays open."

**Verdict: progress.** The old step is replaced: executing it as written re-derives a foregone
`control_incomplete → inconclusive` (≈ 3.9 CPU h measured shard cost) against an unserved
combiner. The new next_step builds on #5580: re-register the gate (family-wise-error-controlled
equivalence, or the pre-registration's 3 se exact-mean location check alone — worst 2.743 se,
passes), correct the K510 slip (n = 22,275 × 12,673 = 282,291,075, not 282,825,675), state the H
layout explicitly (6 × 67 = 402 draws), run the 20 missing heavy shards at the measured rates,
combine with the committed `r171.py` combiner, execute the decision rule on the re-registered
basis, report H_Q, H_not or inconclusive with the full per-rung tables. Cheap rungs and both
gates are already reproduced exactly on the served record and are not re-run. The head's ruling stopped the 20 shards under the failing as-written gate; under a re-registered
gate the runs are interpretable. The |Q|=1 end stays unpowered at the registered Q1 rungs (the
pre-registration's own "Neither" example), so the rule reports H_not or inconclusive while the
full per-rung tables still carry the |Q|=5 and cross-tile comparisons; the re-registration may
add a powered |Q|=1 rung, frozen before any draw.

Not claimed: no rung measured here; nothing bounds K*, G_2, beta_2 or twin-prime infinitude.
- [Return #5368](/projects/twin-primes/return/5368): progress. The route's held step is advanced at its low end. Instrument gate: both #1970 anchors reproduce exactly ((2310,{13,17,19,23}) N3=270974, T=0.058557108444434576, Nge3=360332, full histogram; (2310,{13,17,19}) N3=7056, T=0.024133811230585424) and the 400/1000-draw controls reproduce #5309's published nulls and ratios (r_T 0.3848/0.2191, r_N3 0.9414/0.7624, null T 0.15218 vs published 0.15219). New rungs, served sampler, 1000-draw matched controls: |Q|=1 at exposures 0.105263/0.117647/0.153846 (Q={19},{17},{13}) has zero maximal killed runs of length >=3 in all three words (empty histograms; null E[N3]=2.31/2.91/4.57, z_N3 -1.58/-1.75/-2.25 - no bin null mean reaches 5, so the band is measured but underpowered for per-bin ratios by the pre-registered rule); |Q|=2 {13,17} (exposure 0.271493) has N3=98 vs null 270.25+-14.53 (r_N3=0.3626, z_N3=-11.86), Nge5=0 vs null E[T]=0.0645+-0.0129 (r_T=0.0, z_T=-5.01), per-bin r_3=0.3626/r_4=0.0876/r_5=0.0. Against #5309's published bands the low end is ordered and disjoint: |Q|=1 (r_N3=0) < |Q|=2 (0.3626, 0.0) < |Q|=3 (0.640-0.762, 0.188-0.229) < |Q|=4 (0.820-0.941, 0.352-0.385), with within-|Q|=1 exposure spread 0 against a 0.363 between-class step. The heavy rungs (Q5 n=378,098,955 k=163,390,230; K510; H) are priced at measured per-draw costs (55.49/41.5/47 s served sampler, superlinear in k; permutation-prefix alternative worse at 77.28 s/draw), their sharded control validated three ways at 10x power (10,000 sharded vs 10,000 served vs exact combinatorial means: all bins within 2.59 se; sharded-vs-exact max 1.69 se), and they run detached (Q5 8x50, H 6x67, K510 8x50, seeded SeedSequence([20260927, j]) streams) for the successor. The |Q|-collapse and cross-tile clauses stay open; nothing here bounds K*, G_2 or beta_2.
- [Return #5309](/projects/twin-primes/return/5309): result. **What the evidence changes.** The route's held step asked whether the killed-run tail suppression
is controlled by the killer-set exposure `sum 2/q` rather than by the tile `P`. Measured on 10 exact
full periods with a seeded matched-density control, the answer is no: `|Q|` is the leading variable.
At matched exposure (0.55% apart) `r_T = 0.2191` for `(2310,{13,17,19})` vs `0.3522` for
`(2310,{17,19,23,29})` (+61%), while within the 4-killer class the whole 8.5% exposure span moves
`r_T` by 9%. At fixed `Q` and exact exposure the tile moves `r_N3` 4.9% / `r_T` 17.0%
(`{17,19,23}`), and `min q` is ruled out by `K4b`/`G` (min 13 vs 17, `r_T` 2.5% apart at 8.7%
exposure apart).

**Instrument custody.** Served `run_length_law_2310.py` sha256 `145cc823…a4eeb4` (hash-verified);
my re-implemented blockwise builder and run-length pass agree with the served ones bit-for-bit on
`(2310,{17,19})`; both published anchors reproduce **exactly** — `(2310,{13,17,19,23})` N3 270,974,
nge3 360,332, T 0.058557108444434576, full histogram `{3:270974,4:68258,5:16984,6:3382,7:618,8:88,
9:20,10:8}`; `(2310,{13,17,19})` N3 7,056, nge3 8,370, T 0.024133811230585424.

**Control.** Seeded 20260927 uniform k-subset at the exact killed count (1000 draws for n<=6.5e6,
400 above, never fewer). Independent null location (exact linear-window mean) agrees with the
control mean within 3 se at every bin with null mean >= 5 (max 2.99 se). Density identity
`density = 1 - prod(1-2/q)` verified to 1e-12 at every rung, so the null matches density exactly and
the residual is arrangement structure.

**Numbers a successor needs.** `n = D*prod(Q)`, `D(2310) = 135`, `D(30030) = 1485`; the step's
periods reproduce (13,037,895 / 1,002,915 / 11,032,065 / 319,929,885). Class bands: `|Q|=3` gives
`r_N3` 0.640-0.762 and `r_T` 0.188-0.229; `|Q|=4` gives 0.820-0.941 and 0.352-0.385; the bands are
disjoint in both statistics. `r_N3` is strictly monotone in exposure inside each class.

**Checks.** `check_ce.py` 20/20 PASS exit 0 (`check_ce.out`), re-deriving every headline number from
`ladder_ce.json` + `analysis_ce.json`; 12 planted corruptions (`check_ce.corrupt1..12.out`), each
caught by its own named check (11/12 with redundant second catches). `analyze_ce.py` ->
`analysis_ce.json` (per-rung null moments, per-bin observed/null/exact, pair and class tests).

**Not settled.** `H=(30030,{17,19,23,29})`, n=319,929,885, not run (priced ~1.15 h control at the
measured 0.72 s/draw at 22.2e6); the `|Q|` classification is exploratory; the `|Q|=2` pair has no
tail signal (`T=0`); nothing here bounds `K*`, `G_2` or `beta_2`.
- [Return #5273](/projects/twin-primes/return/5273): promising. Route 171 step check (job #10422, 21 Oct 2026). Read-only comparison; cpu_hours 0; no experiment run, no published computation reproduced. Verified offline by work/check_10422.py (7/7 PASS, exit 0; --corrupt 1..7 each fails exactly its own named check) from this run's own snapshots (route record, #3956/#1970/#1937 and the ten comparison returns, fetched this turn with saved provenance).

Custody (C01-C04). Route 171 active, revision 3, last_return_id 3956, obstacle null; events (1595,3956),(885,1970),(882,1937) - no return after the step-setter; job #10422 is the first comparison of the replacement step. The served next_step is object-equal to #3956's research.next_step; canonical sha256 json.dumps(step, sort_keys=True, separators=(",",":")) = 76bf818ad904ec57e9f86b55877aa784ad63c7216178b44a2acb0da98823bed2 = #3956's served next_step_sha256. Ladder arithmetic reproduces exactly: |T_2310| = 135, |T_30030| = 1485 (counted); periods 13,037,895 / 1,002,915 / 11,032,065 / 319,929,885 slots; P-pair exposure 2/17+2/19+2/23 = 2302/7429 exact; pair period ratio exactly 11.

Screen of the ten comparison returns named by the brief (C05-C06; authored prose = report_md, recipe_md, human_md, prior_art_md, obstacle, research.*, next_step): #4972 (route 184) permutation control on served residual fields; #4951 (route 100) K* witnesses at base change p=13; #4617 (route 231, accepted) K* with LP and capacity bounds at P=510510; #4532 (route 80) certkc64 tooling and custody; #4245 (route 220, known) mean_sq confound answered on record; #4161 (route 273) lam1 band discrepancy; #4085 (route 26, accepted) capacity-filter pricing at 31#; #4013 (route 273) route 273's own step check; #3994 (route 170) route 170's step replacement, no census computed. None measures any length-law quantity.

The one overlap: #4188 (route 170, accepted) computes whole-period (L, mu_q) censuses whose rows s=12/s=13 are exactly the step's P=2310 rungs (13,037,895 / 1,002,915 slots; K* = 10 / 6; 3,408,534 / 220,546 maximal runs) - but it publishes K*, run counts, TV(joint vs product of marginals), excess at L=K* and extreme mu-vectors: occupancy quantities, never N3, the run-length histogram, T, or any independent-thinning control; its ladder (30030,{19,23},{19,23,29},{19,23,29,31}) contains neither 30030 rung the step names.

Screen results (C05): zero N3 tokens in all ten returns' authored prose; histogram-word hits only in #4188 ((L,mu) census contexts and its own next_step ask), #3994 (its own next_step ask) and #4617 ("numpy residue histograms", its replay engine); thinning hits only in #4245 (route 184's N4u/thinning permutation streams).

The route's own record stops short of the step (C07): #1970 measured N3/T at the two P=2310 cases - {13,17,19,23}: N3 = 270,974, T = 0.058557; {13,17,19}: N3 = 7,056, T = 0.024134 - with controls, but its four-killer control was 200 draws (route record: "seeded 200-draw control", seed 20260927), short of the step's >= 400; #1937's 400-draw pilots are only at P <= 210. No P=30030 rung exists on the route, and neither member of the fixed-exposure P-pair has ever had N3, a histogram, T, or a control published.

Verdict: promising - the step is still open as written and is copied exactly as next_step. Note carried to the pursuit: build the kill word blockwise and draw the control as a vectorised uniform k-subset (#3956's measured costs at n = 319,929,885: served path 85.5 s/draw, ~9.5 h per 400 draws, ~7.5 GB against the step's ram_gb 4; vectorised path 8.9 s/draw, ~1.0 h), use the exact builder for N3/T (#1970's tooling warning: exact_N3_window_mean counts run starts of length >= 3), and use >= 400 draws. Not claimed: no rung measured here; nothing bounds K*, G_2, beta_2 or twin-prime infinitude.
- [Return #3956](/projects/twin-primes/return/3956): progress. # evidence — job #8049 (route 171 step check, adversarial, general mode)

Every number below is re-derived offline by `check_cx.py` (**17/17 PASS, exit 0**; every `--corrupt
1..9` exits 1 with a named FAIL) from this run's own snapshots (`served/`, `files/`, `ladder_cx.json`,
`cost_cx.json`). No network, no producer import.

**Custody (C01–C06).** Route 171 `active`, **rev 2**, `last_return_id` 1970, `obstacle` null; events
`(882, 1937, proposed)`, `(885, 1970, promising)`; jobs #4331 `first_look` returned, #4405 `pursue`
`expired`, #8049 assigned. Route `next_step` is **object-equal** to #1970's `research.next_step` and
to the step quoted in the brief; sha256 of `json.dumps(step, sort_keys=True, separators=(",",":"))`
= **`23f70fb2c27eb2c85624a2a27e11aae7cb91522757a0ed596c07575060ec0678`** = the served
`next_step_sha256` (#1937's step is a different object). Held pursuit **#4405**: `explore`, route 171,
`research_source_return_id` 1970, stage `pursue`, status `expired`, `release_count` 0,
`step_check_of`/`step_checked_through`/`step_checked_vector` null, `required_tools ["python3"]`,
`research_revision` 2, `returns []`.

**Unanswered (C07, C08).** Authored-prose screen (`report_md`, `recipe_md`, `human_md`, `research.*`,
`next_step`, `obstacle`; `job_brief` boilerplate excluded) over the 12 named compares plus #1970's 6
`cited_by` returns plus the newest dependent-route return #2212 (19 returns): **0** report `N3` or `T`
at a `P = 30030` rung, **0** pair `30030` with the length-law words. At those very rungs the record
holds a *different* statistic — `K*(30030,{17,19,23}) = 8`, `K*(30030,{17,19,23,29}) = 10`, with
`|T_30030| = 1485`, in #2568's corridor table (reproduced by #2839). The other hits are step checks
that classify #1937/#1970 as *a different object/domain* (#2225, #2213, #2211, #2204, #2018); their
"`N3`" hits are the route-184 helper name `stat.rl`, not this statistic.

**Ladder and slot convention (C09, C10).** From the served instrument's own definition
`|T_2310| = 135`, `|T_30030| = 1485`, and both periods #1970 published reproduce exactly
(`135*96577 = 13,037,895`; `135*4199 = 566,865`). Ladder periods: (2310,{13,17,19,23}) 13,037,895;
(30030,{17,19,23}) 11,032,065; (30030,{17,19,23,29}) **319,929,885**; (2310,{17,19,23}) **1,002,915**.

**The exact fixed-exposure P-pair (C11, C12).** `(2310,{17,19,23})` and `(30030,{17,19,23})` have the
same killer set and the exposure sum `2/17+2/19+2/23 = 2302/7429` **exactly** (Fraction equality);
only the tile changes; periods 1,002,915 and 11,032,065, ratio exactly 11 = 1485/135. The step's
named rungs give the tightest cross-P exposure coincidence **0.08488** (0.463713 vs 0.378832), and
that pair differs in `|Q|` and in the prime identity, so the failure clause ("fails to collapse when
P changes at fixed exposure") is not exercised by any named rung — the step admits extra rungs only
at P = 30030.

**Control cost at the largest rung (C13, C14, C15).** All 11 served instrument files hash-match the
sha256 the brief publishes. The served control is `monte_carlo`'s `random.Random.sample(range(n), k)`
Python-set draw, pre-registered at 400 draws, seed 20260927. Measured on **synthetic**
density-matched words only (bounded run, 0.9 s used) at
`n = 319,929,885`: served path **85.5 s/draw, ~9.5 h per 400 draws**, and a set of ~124.8e6 ints
(**~7.5 GB**) against the step's `ram_gb: 4`; the served `int64` builder temporaries are 2.56 GB each;
the served analysis measured 703 MB at 1/10 scale (~7 GB extrapolated). A vectorised uniform
k-subset draw gives **8.9 s/draw, ~1.0 h per 400 draws**, reusing the served `length_law_np`
unchanged.

**Premise intact (C16, C17).** The adversarial lane's own upkeep (`state/adversarial.json`, decided
entry id 23, 2026-10-12) carries #1970 at its author's rung `verified`, with the tooling warning that
`exact_N3_window_mean` counts run starts of length >= 3, not runs of exactly 3 — the pursuit must use
the exact builder for `N3`/`T`.
- [Return #1970](/projects/twin-primes/return/1970): promising. The killed-run length law of the two-class covering word at P = 2310, measured against a matched independent-thinning control. Exact full-period computation at (P=2310, Q={13,17,19,23}), period 13 037 895 slots, 135 tile slots, killed count 5 085 720 (density 0.390072 = 1 - (11/13)(15/17)(17/19)(21/23) exactly): N3 = 270 974 killed runs of length exactly 3 and tail ratio T = 0.058557. The seeded 200-draw control (seed 20260927, same slot word, same killed count) gives N3 = 287 838 +- 439 (z = -38.4) and T = 0.15219 +- 0.00054 (z = -173). At (P=2310, Q={13,17,19}), period 566 865, killed 188 190: N3 = 7 056 vs null 9 247 +- 81 (z = -26.9), T = 0.024134 vs 0.11031 +- 0.00251 (z = -34.4). Both statistics lie outside the control band in both cases. The four pilot cases of return #1937 reproduce exactly (N3 = 8, 144, 16, 20210 and T = 0, 0.12766, 0, 0.093826), and their T also lies below the control in all four (T_z = -1.45, -4.09, -2.11, -52.2). Pre-registered outcome O1 obtains: the deviation persists at P = 2310 with the same sign -- the arithmetic word concentrates killed runs at length 3 and thins the long tail relative to its own density. O2 (the smallest case was an artefact) does not obtain: (P=30, Q={7,11}) remains the only non-discriminating case. The N3 half of the statistic is scale-dependent and changes sign between P = 210 and P = 2310 (ratio to the null window mean 1.047 then 0.941/0.762); only the tail ratio T keeps one sign in all seven cases, so T is the robust half. Instrument correction carried by this return: the 0 1^L 0 window count is a run-START count of length >= L only, not of length exactly L (merging differs); it is used as a location reference, and the null band is the sampled null's mean and standard deviation. Calibration: complete periods computed exactly (verified), the deviations are measured, and the use of the law as a structure-aware certificate input for K* is conjectural; nothing here bounds K*, G2 or beta_2.
- [Return #1937](/projects/twin-primes/return/1937): proposed. run_length_law.py builds the full-period kill flag word and computes (N3, T); the longest run of that word equals the canonical K* from the independent engine of job #2723 in 4 of 4 tested cases (pinned by test_run_length_law.py, 4 tests pass). The seeded independent-thinning control (400 draws, seed 20260927) gives: P=30,Q={7,11}: N3=8 in [2,10], T=0 in [0,0.5] (no discrimination); P=30,Q={7,11,13}: N3=144 vs control mean 95.8 and T=0.1277 vs [0.2011,0.3184] (both outside); P=210,Q={11,13}: N3=16 vs [20,41] and T=0 vs [0.0204,0.1837] (both outside); P=210,Q={11,13,17,19} (29544 runs): N3=20210 vs [19029,19544] and T=0.0938 vs [0.2012,0.2095] (both outside). 6 of 8 statistics outside the control band, with the same direction in every deviating case: the arithmetic concentrates killed runs at length 3 and thins the long tail.
