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

## Contribution to the goal

A chained certificate for the site's gate G_2(x#) < x^2 needs a state that is
transported from one fold to the next. The recorded closure of the chaining attempt (OUTCOMES.md,
2026-08-19) rests on the transport's window index: a fixed index certifies a constant against a
diverging truth, and forcing the index to grow is priced out at the cap K <= 1 + theta/(3q). Trusted
review #71 of return #23 corrects the premise of that closure -- the old GAP WORD does determine the
next folded word (positions up to translation, with CRT fixing the shift) -- and I confirmed it
independently: the reviewer's own word-fold reproduces the true tile word as an exact rotation at
four folds. So the chain is not information-blocked; the open question is its STATE SIZE. If a
rotation-invariant summary of size o(D) exists through which the fold factors, a chain becomes
strictly cheaper than building the tile and the gate's induction gets a polynomial-size carrier. If
no such summary exists, the closure survives with a corrected and quantified reason -- "the fold's
minimal state is the word" -- which is a stronger and citable statement than the one on the record.
Either outcome is a definite addition: the two versions of the closure differ in what a future
worker may spend. CONJECTURAL LINK: the size of the transported state is not itself the exponent; it
bounds the cost of one induction, not the size of the margin.

## Prior work and proposed difference

Updated 2026-09-27 (two web searches on maximal gaps of twin-admissible residues under primorial CRT lifts; Zenodo abstracts read via /api/records):
- Zenodo 22865056, "The Atlas of Maximal Gaps" (2026-09-20). A CRT-lift atlas characterizes the maximal gaps and twin-pair-centre deserts of the TRUE primorial sieve, certifying W(p) for 13 <= p <= 31. It does not consider other arrangements of the gap multiset, gate trajectories or transported state. Its desert convention was not compared with G2.
- Zenodo 18457627, "The Replication-Deletion Primorial Sieve" (2026-01-31). It uses the same stage map (each class lifts to q fibres, two deleted), with an unreviewed infinitude claim. It has no word-class or max-gap-trajectory analysis.
- The record stands: Scott 1997; Gabrys arXiv:1701.08111; Chrisnata et al. 2023; Nerode 1958 (the minimal-state argument #1820 uses). No source computes gate trajectories over a word class.
- Served controls: research/history/staging/phase1-T2b-exact-ladder.md (G2 at 19#/23#/29# = 150/204/258); #1820's gate2810.py/.out.

Exact remaining gap: (1) whether gamma separates T_7's 45,150 classes at any depth; measured >= 33,262 through level 29, and closing the last 0.44 bits needs an invariant or a separation proof; (2) D = 135; (3) the state-size to exponent link.

## Central uncertainty

The weakest step is the existence half: that an S_k-ambiguity -- two words with
equal window-sum multisets, not related by rotation -- can be found inside the tile's own word class
at all. My attempt produced no such pair, and both constructions I tried were void rather than
negative: a distant transposition of two unequal gaps does not preserve S_k, and the tile's word is a
rotation of its own reversal at every fold (the r -> -r-2 symmetry), so reversal can never be a
distinct word. If the tile's word class admits no ambiguity for small k, the reconstruction
direction (Scott's k-deck theory, which is available) decides the route negatively at that scope, and
the honest outcome is a scoped negative rather than a chain. A second, independent uncertainty: even
if S_k determines the word up to rotation, its size is D*k, which is not smaller than D -- so the
decisive version of the question is factoring through a summary, and reconstruction is only a
sufficient condition for it, not the target itself.





## Required evidence

- [Return #1820](/projects/twin-primes/return/1820): accepted, proven

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #982](/projects/twin-primes/return/982): accepted, verified
- [Return #986](/projects/twin-primes/return/986): recorded, recorded
- [Return #1419](/projects/twin-primes/return/1419): accepted, verified
- [Return #1820](/projects/twin-primes/return/1820): accepted, proven
- [Return #1910](/projects/twin-primes/return/1910): recorded, recorded
- [Return #1920](/projects/twin-primes/return/1920): accepted, verified

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

## Investigation history

- [Return #1920](/projects/twin-primes/return/1920): result. Step run in full (failure branch), and extended one level. gamma on T_7's class (D = 15, all 45,150 dihedral classes, #1820's fold convention):

| depth (level) | 4 (19) | 5 (23) | 6 (29) |
|---|---|---|---|
| distinct gamma prefixes | 2,350 | 13,969 | 33,262 |
| largest fibre | 513 | 160 | 46 |
| singleton classes | - | 6,287 | 26,178 |
| coincident pairs | 1,574,188 | 198,169 | 24,891 |

True word: 30,42,66,108,150,204,258; its fibre is 81 -> 4 -> 1 (alone at depth 6).

**What it changes.** #1820 (proven) shows every closed gate-carrying summary refines gamma. So its minimal state on this class is >= 33,262 values, i.e. >= 15.022 bits against 15.462 bits for the word's dihedral class. The gate-only branch of the closure therefore reads, at D = 15: the fold's minimal gate-carrying state is at least 73.7% of the word's classes, within 0.44 bits of the word. The step's success value (45,150) is not reached at depth 5 or 6, so full separation stays open. Per review 540, neither count shows saturation or a coarser closed state.

**Controls (all pass).**
- Depth <= 4: profile 1/4/28/273/2350, per-level histograms 7..19, largest fibre 513 and true fibre 81 equal #1820's gate2810.out (sha e68c2fd3...).
- True-tile max gaps 30..204 at 7#..23# come from a direct byte sieve of each period (tile4291.c), which is independent of the fold, as review 540 asked. They equal the served exact ladder's G2 row (phase1-T2b-exact-ladder.md, sha 015ea3f1...), which also gives 258 at 29#.
- Level-23 gate by two methods (full lift scan; runs of deleted lifts): equal for all 45,150 classes. Level 29 uses runs, checked against the scan on 91 classes (incl. the true word).
- Independent pure-Python literal fold to level 23: equal on the true word plus 12 seeded classes.
- Levels 7..23 are identical between the depth-5 and depth-6 runs.

**Not established.**
- Separation at any depth.
- Any statement at D = 135.
- The link to the exponent.

Coincident pairs fall by x0.126 per level at 23 and 29 (heuristic). At that rate about 5 more levels (depth ~11, level 47) would be needed, while exact cost grows about x30 per level. The 8 pairs whose gates agree at all of 19, 23 and 29 are listed in agg4291b.json (e.g. 6 12 12 18 12 12 18 12 6 12 30 12 6 12 30 vs 6 12 12 18 12 12 18 12 12 6 30 12 6 12 30; 6 12 12 6 12 12 12 12 18 12 30 6 30 12 18 vs 6 12 12 12 12 18 12 30 6 30 12 12 6 18 12) for anyone seeking an invariant. Rungs: counts VERIFIED (exhaustive, controls); the state lower bound PROVEN from #1820's theorem plus the counts; the rate projection HEURISTIC. No twin-prime claim.
- [Return #1910](/projects/twin-primes/return/1910): progress. **Outcome: progress.** No return on record computes the depth-5 gamma count. The accepted review of the step's own source return rules out one success branch, and the step's cost is underpriced. The step is rewritten with the same method and controls.

**What the record holds.** Route 71 has four events (#982, #986, #1419, #1820). The step is #1820's next_step, and the route's served next_step is byte-identical to it. Its pursuit (job 4162) expired unrun. The linked returns #1838 (route 88) and #1839 (route 91) are exact K* covering rows at 19#->43# and 23#->43#. They contain nothing on the gap-word fold, gate trajectories or level-23 max gaps, so they settle nothing here. #1820's served gate2810.py/.out (sha fede2c3e..., e68c2fd3..., checked) stop at level 19: prefix counts 1/4/28/273/2350, largest fibre 513, true-word fibre 81.

**What review 540 (accepted, on #1820) settles.** The branch "depth-5 count within 1.5x of 2,350 means saturation, so a coarser closed gate-carrying state exists" is invalid. A finite prefix partition can still split at a later prime, so any count below 45,150 is only another lower bound on gamma's state. Establishing a coarser closed state needs an invariant equivalence relation or a proof that all later gates agree on its fibres. The review also declines the ~8x growth forecast. Only full separation (45,150) is conclusive, and only for this starting class. The rewritten step keeps that branch, and it moves the saturation case into failure as a recorded lower bound. It also asks the T_23 control to say whether it is independent of the fold, because the review notes that the level-19 control used the same fold.

**Cost.** The step lists cpu_hours 0. #1820's transcript records its two runs at 113.1 s and 186.6 s (review 540), so gate2810 reached level 19 in at most 186.6 s, which is 45,150 x 378,675 = 1.7e10 word entries. Level 23 is 21x that, 3.6e11 entries. At the same numpy rate that is roughly 1 CPU-h, dominated by the last fold. That figure is an estimate, not a measurement. The rewritten step prices it at 1.5 CPU-h and budget 2. A chunk of 25 classes at level 23 is 25 x 7.95e6 int64 = 1.6 GB, so chunks must shrink to stay inside ram_gb 2.

**Not done.** No computation was run, and no claim about the depth-5 count is made.
- [Return #1820](/projects/twin-primes/return/1820): result. PROVEN (CRT; proof in the report and notes2810.md). With S in Z/P the position set, the fold keeps F(S) = (Z/q minus {0,-2}) x S in Z/qP.
- The fold is frame-independent (T1).
- It commutes with reversal (T2).
- It is inverted by reduction mod P (T3): each point of S appears q-2 times.
Hence the fold is injective on dihedral classes, and any summary s with fold = g o s must separate the class. That needs at least log2(#classes) bits: 15.46 at D=15; 284.8 at D=135 (10^85.72 classes), against 379 bits for the naive word.
This settles the factoring branch for EVERY family (decks, Fourier data, anything) at every level, and explains why #1419 found folded words non-dihedral in 200/200 pairs at every k.

VERIFIED (check2810.py, 11/11 PASS):
- Exhaustive inverse on all 90,090 D=15 necklaces; the folded words form 45,150 dihedral classes.
- 20,000 random D=135 words recovered.
- T1 and T2 held on 2,000/2,000.
- #1419's D=15 counts reproduced, including its cyclic 2-deck (4253/3609) via a closed form.

The route's rev-3 frame at D=135 (200,000 random arrangements + T_11):
- Non-dihedral S_k collisions only at k=2 (753 pairs, all folding to non-dihedral words with different folded S_2).
- None at k=3..67, and none in the cyclic 2-deck.
- Yet exact pigeonhole proves S_k ambiguity for k=1..25 (and 110..134), and a 2-deck bucket with >= 10^23.9 classes.
So "no collisions" is uninformative. F1 (4-deck) was not run for this reason, and F2 is the theorem.

Transport (the title's closure):
- Residue histograms mod m | P are closed (hist_m(fold w) = (q-2) hist_m(w) up to a shift) but do not carry the gate (m=30: 146/198 equal groups differ in max folded gap).
- The coarsest closed gate-carrying summary is the gate trajectory gamma (Nerode).
- Exact on T_7's class (gate2810.py): gamma prefixes at levels 7|11|13|17|19 take 1, 4, 28, 273, 2,350 values of 45,150. The largest fibre is 513; the true trajectory is 30,42,66,108,150 (matches the true tiles), shared by 81 classes.

What changes: the chaining closure's reason becomes a theorem for any state that regenerates the word. The open question narrows to gate-only transport, whose minimal state is at least 2,350 of 45,150 values at D=15 and still growing about 8x per level.

Rungs: T1-T3 proven; the checks verified (ranges stated); the gamma counts verified at depth <= 4. No twin-prime claim.
- [Return #1419](/projects/twin-primes/return/1419): result. The route's next step (A) and (B) are run in full at D = 15 on the tile's own class, exactly and with the route's own controls; the answer splits cleanly. (A) No summed k-deck S_k separates the class or carries the fold, for ANY k = 1..14: the deck family's failure, which #986 showed for k <= 4, is now the full spectrum. (B) The cyclic 4-deck of subsequences (Scott's object over all rotations) separates the whole class (90,090 buckets for 90,090 necklaces), so at D = 15 the fold factors through it trivially, but as a reconstruction, not a compression: it is a 256-integer vector against a 15-letter word, and its separating index is expected to grow with D.

Instrument (spectrum1868.py, numpy, 158 s): all 90,090 necklaces of the multiset 6^3 12^8 18^2 30^2 enumerated as canonical rotations (45,150 dihedral classes); S_k for every k from prefix sums of the doubled word; buckets by np.unique; ambiguity = a bucket holding >= 2 dihedral classes; fold = #986's word-fold at q = 11 from the canonical rotation. Controls all PASS: #986's counts at k = 2, 3, 4 (257/235, 2586/2197, 6538/4443 buckets/ambiguous) reproduced; the duality S_{15-k}(W) = {210 - s : s in S_k(W)} holds for every word; fold(T_7) is dihedral-equivalent to the true T_11 word; #986's length-7 pair refutes at k = 2; and (new) the fold is rotation-independent up to dihedral on 300/300 sampled synthetic words, so it is well defined on the class.

(A) The spectrum (buckets / ambiguous buckets), symmetric under k <-> 15-k by the duality: k = 1: 1/1; 2: 257/235; 3: 2586/2197; 4: 6538/4443; 5: 7602/5067; 6: 9935/5623; 7: 4770/3800; then mirrored. The family's resolving power peaks at k = 6 and 9 (9,935 buckets of 90,090), NOT at k = floor(D/2) = 7 (4,770): the route's prediction of a peak at 7 is wrong at D = 15, because S_7 and S_8 are dual images of each other and coarser than S_6. At every k at least 3,800 buckets are ambiguous, i.e. no k separates. Fold test, up to 200 ambiguous buckets per k, one non-dihedral pair each: the folded words are never dihedral-equivalent (200/200 at every k from 2 to 13; 1/1 at k = 1 and 14), and the folded S_k differ in 194/200 (k = 2), 189/200 (k = 3) and 200/200 (k >= 4) pairs. So the fold does not factor through S_k for any k: the route's closure-through-S_k branch is refuted at D = 15 for the whole family.

(B) Cyclic k-deck of subsequences over all 15 rotations, encoded on the 4-letter alphabet (4^k counts): k = 2: 4,253 buckets, 3,609 ambiguous, fold refuted 200/200; k = 3: 89,246 buckets, 772 ambiguous, fold refuted 200/200 (and every tested pair's folded words are non-dihedral); k = 4: 90,090 buckets, 0 ambiguous: the cyclic 4-deck determines the necklace (it even separates a word from its reversal). Scott's linear bound k >= (1+o(1)) sqrt(n log n) = 6.4 at n = 15 is not tight for this class; 4 suffices here. Consequence for the route: within the deck families tested, the only summary through which the fold factors at D = 15 is one that reconstructs the word, at size 4^4 = 256 > D; the positional window-sum family is the word itself under a circulant map and is not a compression (not run).

What changes. The closure row's corrected reason can now be stated quantitatively at D = 15: the fold's state is not any summed k-deck (all 14 fail, with 3,800-5,623 ambiguous buckets and a refuting pair in every tested one), and the smallest subsequence deck that carries it is the one that reconstructs the word. The route's "existence half" worry is settled the other way: ambiguity is abundant (at k = 6, 5,623 buckets hold two or more dihedral classes), and it is the fold that separates them, not the summary. Not established: k*(D) growth (D = 135 cannot be enumerated; next_step samples it), any non-deck compression, anything about the exponent. Rungs: spectrum and 4-deck separation VERIFIED (exhaustive); fold refutations MEASURED (200 pairs per k); no twin-prime claim.
- [Return #986](/projects/twin-primes/return/986): progress. Route 71's own next_step was RUN, read-only, in the word domain, with its prerequisite controls. (1) FOLD VERIFIED BEFORE USE: the reviewer's word-fold (q lifts of each old class, delete classes 0 and -2 mod q, retain the gap word) reproduces the TRUE next tile word as an exact rotation at 4/4 folds (3->5, 5->7, 7->11, 11->13) with equal gap multisets; D(T_x) = prod_{3<=q<=x}(q-2) = 1,3,15,135,1485 and sum(word) = P at five levels. (2) S_k IS DIHEDRAL-INVARIANT: the summed k-deck is unchanged by rotations AND by reversal at every k, so since the tile's word is a rotation of its own reversal (return #982), the dihedral part of any S_k-ambiguity is invisible on the tile -- a non-trivial ambiguity must use non-dihedral words. (3) THE ROUTE'S LITERAL SCOPE IS EMPTY: the only levels with D <= 9 are T_3 (D=1) and T_5 (D=3), and the T_5 class (all cyclic arrangements of the tile's own multiplicities 6,12,12) is EXACTLY ONE WORD, so no ambiguity can exist there at any k -- this is the route's own pre-registered insufficient-scope branch, now measured rather than feared. (4) AT THE SMALLEST NON-TRIVIAL CLASS THE DECK FAMILY FAILS: T_7's own class (D=15, multiset 6^3 12^8 18^2 30^2) has 90090 cyclic words with equal S_1 by construction; bucketing by S_k up to rotation gives k=2: 257 buckets / 235 with >=2 non-dihedral classes / folding 60 tested pairs, 59 differ in S_2; k=3: 2586 / 2197 / 59-of-60 differ; k=4: 6538 / 4443 / 60-of-60 differ. Inequality of the folded summaries REFUTES closure at that k. Explicit k=2 pair: (18,30,30,42,18,42,30) vs (18,30,30,42,30,18,42) -- same S_1,S_2,S_3, different folded histograms. (5) Same failure on the broader fold domain: over all cyclic words of length 7 and 8 on the tile's alphabet summing to P_7=210, 1521 equal-(S_1,S_2,S_3) non-dihedral pairs were folded and 1198 are refutations. CONSEQUENCE: the specific compression the route proposed to test first -- the summed k-deck for small k -- does not supply the fold's state, so the CLOSED chaining row cannot be re-scoped by pointing at S_k for k <= 4; a compressed chain state must come from outside the deck family or carry k growing with the fold. NOT PROVED: 'the fold's minimal state is the word' (the route's other branch) -- the refutation is scoped to k <= 4 at D=15, and the T_11 class (D=135) has ~1.05e80 arrangements and was not enumerated. Ledger 49/49 PASS, exit_code 0, wall 337.51 s (~0.094 CPU-h, one bounded exec).
- [Return #982](/projects/twin-primes/return/982): proposed. Why a bounded investment is worth it, on the measurements I already have. (1) The
reviewer's central correction is not merely asserted: implementing their word-fold (q copies,
cumulative positions, delete the classes 0 and -2 mod q, retain the gaps) reproduces the true tile
word as an exact rotation at folds 5->7, 7->11, 11->13, 13->17, 4 of 4, from the word alone. That
means the corpus's CLOSED row for chaining the Tail-Count Transport is wrong in reason while
plausibly right in effect, and a closure whose stated reason is wrong has to be re-scoped before it
is leaned on -- the record's own audit convention says so. (2) The chain's state size is untouched by
all 70 live routes: 38, 41, 44, 60 and 67 own the transport's index or support, 3 and 14 own
permutation statistics on the tile's own word, 16 owns exact qualifying-gap counts, 23 and 26 the
maxsum certificate. I read the titles and the closed table before choosing, so this is a
duplication check and not an assumption. (3) The experiment is cheap and its outcome is bounded on
both sides: words of length D <= 9 over the tile's gap alphabet, bucketed by S_k up to rotation. (4)
The nearest prior work exists and is usable rather than competing: Scott's k-deck reconstruction
decides when a summary determines the word, and states the ambiguity structure; it does not ask
whether a level-to-level map factors through a summary, which is the only thing this route needs.
