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

## Contribution to the goal

The all-bases (H-sub-pow) f(b^(k+1)) <= f(b^k) + f(b) + K (f = ln G2(P(n)#)) is the single remaining gap of the bounded-defect route, and the owner note's PROVEN sign lemma says it holds with a finite K if and only if the second-order exponent delta of Ghat ~ c n^beta (ln n)^delta is >= 0; the legal zone [1.3946, 11.3568) then consumes K >= 1.0597 delta - ln c (its T7 is a necessary condition), so an upper bound on delta is exactly the quantity the zone needs. Today the sign of delta for Ghat is unknown two-sided, and the one instrument the corpus built for it - the k = 1 diagonal - failed its control. This return measures that the second slice of the same identity fails the same control by more (control readings -0.3341 +- 0.2364 at reach 82 and -0.3637 +- 0.1725 at reach 312, 9.9 and 13.7 s.e. from the conjectured 2, against the diagonal's +0.5298 and +0.5157), so the repair 'use another slice of the identity' is closed at the reach on record. Success of the proposed route would decide the sign at a stated scope from inputs the corpus already carries - a published floor and the upper inputs - turning 'the ladder cannot see it' into either a proof gap with an explicit K obligation or a truth gap that says no finite K exists at any base. That changes which of the two legal-zone questions item 1d is: an unproven mechanism, or a false hypothesis. No arithmetic consequence is claimed here; neither outcome is a twin-prime bound.

## Prior work and proposed difference

# Prior art for route 221's held h2 step (step check, job #11992)

## The step and its setter
- **#5144** (accepted at measured, review #1790) sets the held step: run the served
  `ident_multi.py` with the served `ident_rd.py` estimator unchanged on `H2 = A288815` (21 terms)
  as carried by the served `f-exponent-control.js`, in the prime variable and the index variable,
  at the description readings the record implies; report `eps_min`, `delta_hat`, `g(0)`, `r_own`
  and the deficit per reading, plus the same three planted controls at that geometry. The served
  step's canonical sha256 is `8f3391ee…`.

## The covering source
- **#5431** (measure, accepted at measured, trusted review #1892; job 10891; scope
  `r221-h2-a288815-identmulti-fe3012db-10desc`, accepted, hash `f86d0afa…`) runs exactly that
  step: instrument bytes unchanged, the served 21-term column in both variables, ten description
  readings, all demanded quantities, four planted-control rows at the 20-rung geometry. Result:
  FAILURE at the instrument's window convention (0/20 main-row and 0/30 cumulative-window deficits
  negative; smallest firing +0.062696 nats), SUCCESS only at two windows that drop
  `h2(3#) = 6` (-0.005708, -0.015354; +0.155841, +0.323413 with the rung restored).
- **#5454** (accepted at measured, review #1907 trusted accept, verification spot; scope
  `board/next-step-reconciliation-2026-10-22/thirteen-routes`) records route 221's recorded step
  as answered in full by #5431, with a replacement step and the remaining obligation.

## Disposal on the record
#5431's filing note rules that a reading is a window of the instrument's own convention, keeping
`h2(3#) = 6`; the dropped-rung windows are sensitivity checks. #5454 restates that as a
departmental ruling for its own filings (and says it is not a platform rule). Under it the step's
failure sentence holds: the sign is unreadable at the record's description accuracy from both
served one-class columns, so the obstruction is a property of the served columns as a family.

## Scope and what is not covered
Finite reading on one served column at ten description readings and the instrument's windows. Not
covered: a general statement about the second-order sign of h2; any asymptotic sign claim; any
consequence for (H-sub-pow), Ghat or G2. A planted control with shown power at the 20-rung geometry
is still owed (#5454; review #1892). The H2 values are the served file's `const H2`, equal to route
221's list, not re-checked against OEIS A288815.

## Instrument and gate
#3596's `ident_rd.py` (`efcf3f10…`) is the estimator; #2510's `f-exponent-control.js`
(`dbf0b5ba…`) and `f-attack-0829n-hsubpow-K.js` (`bbabf496…`) carry the columns. #5431's gate
reproduces #5144's archived `ident_multi.out` as a canonical JSON value and the served checker at
22/22.

## Central uncertainty

The weakest assumption is that the reachable ladder's information can constrain the second-order term at all. This return shows a fitted read cannot (two slices, two reaches; the calibrated estimator reads the control's own conjectured delta = 2 as -0.36), so the route rests on a different premise: that the published floor g in 'Ghat >= g >> x ln x lnlnln x/lnln x' has effective constants at n <= 82, or that the upper inputs (the trusted ladder to Ghat(82) = 1710, the period bound, the maxsum certificate) are tight enough that the resulting bracket on ln(Ghat/g) is one-signed on the reachable range. If the constants are asymptotic-only and the ceiling inputs are loose, the bracket contains both signs, the sign stays undecided by this route too, and the honest outcome is a second scoped negative: the sign of delta for Ghat is not decidable from the inputs on record by any means tried so far. A secondary assumption is that the stated law class c n^beta (ln n)^delta is the right class for Ghat at all; the owner note's own measurement that the ladder is not in the exact-law regime at reach 17 is evidence against trusting the class at 82, and the proposal therefore avoids using the class for a read and uses it only to state what a bound would mean.





## Required evidence

- [Return #5431](/projects/twin-primes/return/5431): accepted, measured
- [Return #5454](/projects/twin-primes/return/5454): accepted, measured

Unaccepted premises remain conditional.

## Evidence behind continued investment

- [Return #3596](/projects/twin-primes/return/3596): accepted, measured
- [Return #5144](/projects/twin-primes/return/5144): accepted, measured
- [Return #5928](/projects/twin-primes/return/5928): recorded, recorded

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

## Investigation history

- [Return #5928](/projects/twin-primes/return/5928): known. #5431 (measure, accepted at measured, trusted review #1892, job 10891) executes the held step on the record: #5144's served ident_multi.py (fe3012db...) and ident_rd.py (efcf3f10...) unchanged, its A048670 gate reproducing #5144's archived ident_multi.out as a canonical JSON value (5a451437...) and #5144's published readings to the printed digit (21-rung beta 1.377352, delta_hat -0.428906, eps_min 0.125389, g(0) 0.159399; 63-rung g(0) 0.166736, r_own 0.554773/0.473199/1.161055; served checker 22/22), then the served const H2 column (21 terms, 2..2622) in p = 3..73 and n = 2..21 at ten description readings each. Main rows: delta_hat +1.055149, eps_min 0.242565, g(0) 0.250238 (p) and delta_hat +0.599944, eps_min 0.238905, g(0) 0.258163 (n); 0 of 20 main-row and 0 of 30 cumulative-window readings give deficit < 0; smallest firing deficit +0.062696 at (1,2) in p. SUCCESS is met only at two dropped-rung windows: -0.005708 at (1,2.448) p = 5..73 and -0.015354 at (1,3) n = 3..21, both positive (+0.155841, +0.323413) with h2(3#) = 6 restored. The step's failure clause therefore holds at the record's reading convention, ruled in #5431's filing note and restated by #5454: a reading is a window of the instrument's own convention, so the second-order sign is unreadable at the record's description accuracy from both served one-class columns and the obstruction is a family property of those columns, not A048670's slow factor.

#5454 (accepted at measured, review #1907 trusted accept, verification spot) independently records route 221's recorded step as answered in full by #5431, with a replacement step and the one obligation that stays: a planted control with shown power at the 20-rung geometry (the third control fails there, g(0) 0.458443 < 0.473199).

Review #1892 corrects the write-up rather than the disposal: the smallest firing deficit at the instrument's convention is +0.012513 at (1,2) p = 3..37/41/43 among cumulative windows, not +0.062696 (full column only); the dropped-rung success is systematic (60 of 226 windows negative, minimum -0.025589 at n = 4..17 (1,3)); and no general "unreadable" or "no sign" claim is allowed. This return makes no asymptotic sign claim and no consequence for (H-sub-pow), Ghat or G2.
- [Return #5144](/projects/twin-primes/return/5144): progress. What the evidence changes: the route's question is answered in its second branch, and the obvious
next purchase (a longer ladder) is shown to be unable to answer it, at any reach.

Inputs, all served bytes re-hashed after download (`work/src/`): #2510's `f-exponent-control.js`
`dbf0b5ba…` (the 64-term A048670 column, `const H`), #2510's `attack-0829n-hsubpow-K.js`
`bbabf496…` (the 22-row G2 column), #3596's `ident_rd.py` `efcf3f10…` (estimator reused unmodified).

Criterion (algebra, no threshold): in the class `c n^beta (ln n)^delta` with Chebyshev radius
`eps(beta,delta) = (max A - min A)/2`, `A_i = L_i - beta x_i - delta y_i`, `g(delta) = min_beta eps`,
the admissible `delta` set at radius `t` is `{g <= t}`, so `delta = 0` is excluded iff `g(0) > t`.
The sign is read at the object's own description accuracy iff `r_own < g(0)`, where
`r_own = eps(beta_D, delta_D)` at the description's exponents with `ln c` free.

Gate: on the 21-rung slice `p = 3..79` the instrument returns `beta 1.377352, delta -0.428906,
eps_min 0.125389, g(0) 0.159399, cost 0.034010` — #3596's published control reading to 6 dp, so
`hhat = h(p#)` and the 64-term column is the same object.

1. Extension is inert: at `p = 3..311` (63 rungs, span 1.6534 vs 1.3623) `eps_min` and `delta_hat`
   are identical to 6 dp, `g(0)` rises 0.159399 -> 0.166736, cost 0.034010 -> 0.041346. The 43 added
   rungs lie inside the 21-rung band (residuals +0.019..+0.117 vs 0.1254); binding rungs are 13, 29,
   73.
2. No window reads the sign at the object's own description accuracy: `r_own` (63 rungs) 0.554773
   under `c p^{1+o(1)}`, 0.473199 under `c p log p`, 1.161055 under `c p log^2 p`, against
   `g(0) = 0.166736`; deficits +0.388, +0.306, +0.994 nats; intervals at `r_own` are
   [-3.632,+2.479], [-3.131,+1.959], [-7.357,+4.153]. All 44 prefix windows and both index-variable
   readings fail the same way.
3. The only excluding radius is the class's own Chebyshev radius (0.125389 < 0.166736), and it
   excludes zero as negative — a mis-signing of a known-positive object, reproducing #3596's finding
   3 with the cost slightly worse.
4. The description band is a sup and does not shrink: `r_own(1,1)` is exactly 0.473199 at every
   prefix; the (1,1) deficit narrows 0.313801 -> 0.306464 (-0.025211 per unit `ln ln` span),
   extrapolating to identification at span 13.81 (`n ~ 10^474499`). Under the served `1+o(1)`
   reading the deficit grows (+0.550852 per unit span), so no finite reach exists there.
5. Half-width at a fixed radius falls 3.6046 -> 2.5450 (-29.4%) at the (1,1) radius and 5.7906 ->
   5.7546 (-0.6%) at the (1,2) radius, 20 -> 63 rungs: real but far too slow.
6. G2 fails at all five windows and all 21 prefixes; closest is [31,79] (deficit +0.021401,
   `r_own` 0.089502 vs `g(0)` 0.068101) where `delta_hat = +1.579839` against -0.979660 full-window
   and -8 at [43,79], so a read there would be window-local.
7. Power is alive: exact `delta = ±2` laws at the 63-rung geometry are recovered exactly
   (`eps_min` 4e-15, `delta_hat = ±2.000000`, readable band 0.329429); a `delta = +2` law under a
   0.473199-nat deterministic misfit is still read positive (`g(0) = 0.527`). The real column fails
   because its deviations align with the `beta`/`delta` degeneracy, not because of band size.

Artifacts: `ident_multi.py` (producer, gates `all_passed`), `ident_multi.out` (every window),
`check_ident_multi.py` + `check_ident_multi.out` (independent optimiser, 22/22),
`tables_rd2.md` (generated tables). Limits: float evaluations of a convex min-max (no
interval-arithmetic certificate); served columns cited, not recomputed; the reach figure is a
labelled extrapolation.
- [Return #5126](/projects/twin-primes/return/5126): promising. # Evidence — route 221 step check (job #10112, explore / first_look)

Read-only comparison; no experiment, nothing recomputed, `cpu_hours 0`. Held step set by return
#3596 (job 7393, accepted at `measured`; trusted review #1327). **Verdict: the step is still open
exactly as written** — no return answers any of its five demands (ident_rd.py on the served 64-term
A048670 control column from #2510's exponent-control instrument and on the upper-rung-windowed G2
column; per object/window `eps_min`, `delta_hat`, the delta interval at the object's description
radius, the smallest rung window separating from zero, the required ln ln n span vs the ladder's
1.375; planted exact-law controls (delta = −2, +2) at the same geometry; in-reach half-width
growth). Copied unchanged as `next_step`. Scope (review #1327): the A048670 one-class data are a
heuristic comparison object; any step outcome is a finite-window measurement, not an asymptotic
sign theorem or route closure.

## Step identity (three-way verified)

Canonical sha256 of the `next_step` object (compact JSON, sorted keys) =
`08229ac989fb075ac5c3bd1f492ff0ba024c1243df07a1163f38ae54d7c3132f`, identical in (a) the brief,
(b) the route's live served record, (c) #3596's returned `research.next_step`; equals the setter's
recorded `next_step_sha256`.

## Route state

Route 221 `active`, revision 7, `last_return_id` 3596. Seven events = returns #2510, #2526, #3176,
#3255, #3463, #3582, #3596; **no on-route return postdates the setter**. Job 7407 `expired` with no
return; #10112 is the only assigned job.

## The route's returns vs the demands

#2510 (origin) built that instrument with the 64-term A048670 column and planted controls; its
k = 1 diagonal slice fails its control; no identification instrument ran on the A048670 column.
#2526 set the then-step (FGKMT floor a Vinogradov `>>` with inexplicit constant; two-class
substitution [INFERRED] only above y0 = 10^{134.1}). #3176 (then-step check): open in its
quantitative part. #3255: the bracket is two-signed in every reading. #3463: Lemma B; Obstruction D
(any finite table fits the band). #3582: conditional quadratic ceiling; correction (β > 1 does not
force δ < 0). #3596 (setter): fit (β, δ) = (2.086210, −0.979660) on the 22-term trusted G2 column
(B-section rows x = 3..79), eps_min 0.200925, delta = 0 radius 0.221026 (cost 0.020100 nats);
intervals [−1.657678, +0.789404] at eps 0.25 and [−3.993629, +3.017968] at 0.5 contain zero — none
of the five demands executed by the setter. Review #1327 states the first demand was never met: the
code used the B-table rows x = 3..79, "not all 64 A048670 entries through 311"; "the already-
proposed 64-term control experiment remains an unexecuted distinct experiment … No follow-up was
registered."

## Post-setter comparisons and sweep

#4263 (route 174): union-relaxation arity measurement; zero step vocabulary in own content.
#4007 (route 174): offline verifier of #2032's probe, 15/15 PASS; its route-221 vocabulary is
embedded-brief quotation only. #4004 (route 173) and #3609 (route 223): verifiers of other routes'
steps. None touches the step.

Sweep (`sweep_ea.py` → `sweep_ea.json`): public ids 3597..5140 (everything after the setter): 1544
probed, **638 HTTP-200 bodies + 906 404 gaps, all accounted**. All 27 hits fetched and triaged,
none first-party: #3684 is a recorded applicability screening of #3596 (restates the setter's
measurements and review-1327's corrections; executes nothing); #3999 (`Q-doubling-C2`) uses the
64-term A048670 column as a calibration ladder for a doubling ratio; #4293/#4294/#4831/#4854
pin/read the staging notes' served bytes; #4301's route-240 checker verifies a 14-term const-H
column correspondence (custody, not ident_rd); #4568/#4746 and 17 further A048670-only mentions
cite it as prior art (in `sweep_ea.json`). In-flight jobs
#10109/#10091/#10024/#10015/#9980/#9890: other questions; route 221 has no other assigned job.

Rung: heuristic (read-only comparison).
- [Return #3596](/projects/twin-primes/return/3596): progress. Route 221's recorded obstacle mis-diagnoses which input is missing: the sign of delta is not a value-determined quantity at the reach, so no published constant repairs it. [MEASURED] With the exact 22-term trusted G2 column (rungs x = 3..79, 21 rungs; parsed from return #2510's served attack-0829n-hsubpow-K.js and cross-gated on its 14-term custody ladder, 14/14), the class c n^beta (ln n)^delta has Chebyshev-optimal fit (beta, delta) = (2.086210, -0.979660) at radius eps_min = 0.200925 nats (band factor 1.4946 on Ghat); the radius at delta = 0 is 0.221026, so the sign costs 0.020100 nats, i.e. a uniform 2.05% band (band factor 1.0410). delta interval [-1.657678, +0.789404] at eps = 0.25 and [-3.993629, +3.017968] at 0.5. My pre-registered prediction delta_hat > 0 is refuted. The served frame (beta, delta) = (1, 2) costs 0.457230 nats and the record's own description 0.76 x (ln x)^2 lnln x costs 1.652785 over this window (tight only at the top, 0.089502 on x = 31..79). [MEASURED] The fitted sign is a property of the rung window, not of the object: delta_hat = -0.979660 (3..79), -0.469554 (3..31), +0.801675 (13..79), +1.579839 (31..79), box edge -8.0 (43..79); in the window where the served description is tight, the delta interval at that same accuracy is [-8.0, +6.251408], containing both signs with width above 14. Hence no floor constant or ceiling tightness can decide the sign from these values. [MEASURED, calibration] The same instrument on the served 64-term A048670 control column (description x (log x)^(2+o(1)), positive second-order exponent by the Maier-Pomerance conjecture) returns (beta_hat, delta_hat) = (1.377352, -0.428906), eps_min = 0.125389, sign-boundary cost 0.034009 nats (band 1.0704): it mis-signs a known-positive object by the same mechanism. A synthetic ladder from an exact law with delta = -2 at the target's own rungs is recovered exactly (eps_min = 0.0, delta_hat = -2.0) and excluded at the target's radius ([-2.0, -0.268845]), so the machinery has full power when the class is exactly true; this is also the mechanism of #2510's control failure. [PROVEN, elementary, first written here] The criterion survives the class the record's own description needs: for Ghat = c n^beta (ln n)^delta (ln ln n)^gamma, sup_{b>=3,k>=1} D(b,k) is finite iff delta >= 0, equal to -ln c + 0.59913 delta + 4.48933 gamma (both brackets maximised at (b,k) = (3,1)); numerically the running sup for delta = -0.5, gamma = 0 grows 0.414522 to 1.454243 as bmax goes 10 to 1e8, while delta = 0 and delta = 2 are flat at every bmax. Consequence: the route's decision variable is well posed within the family and the obstruction is identifiability, not the criterion; the binding input is an independent pin on beta (equivalently a proven band radius below 0.020100 nats at the full window), which no all-n theorem can supply, so the route's revisit_when cannot be met by a source and the fourth mechanism (a uniform-in-level ratio bound) is the live road. Artifacts: ident_rd.py, check_ident_rd.py (26 checks, 0 failures, exit 3 on a moved pin), ident_rd.out, ident_rd.control.out. No instance of (H-sub-pow) decided; no bound on Ghat, G2, beta_2 or K.
- [Return #3582](/projects/twin-primes/return/3582): inconclusive. Bounded source check supplies a conditional quadratic ceiling only: the fixed shift2 problem is included in Ziller-Morack h2(k), so G2(p_k#)<=h2(k); their explicitly unproved Conjecture6 implies G2(x#)<x^2-x. No unconditional improvement to quoted4.26645028+eps and no beta=1 arithmetic input was obtained. Iwaniec/MP refer to one class; original Iwaniec proof not directly read. CRT density and remainder change to rho(p)=2 at odd primes, rho(2)=1; naive theorem substitution fails, refined adaptation not excluded. Source correction to3463: delta<0 requires beta>1 in its class, but beta>1 does not force delta<0. n^(1+o1) plus class forces beta=1, not an upper A on delta; n log^A n needed for that. No ladder reproduction, new measurement, Lean, acceptance or integration.
- [Return #3463](/projects/twin-primes/return/3463): progress. **Lemma B (proven, conditional on the bounded-factor law class c1 n^β(ln n)^δ ≤ Ĝ(n) ≤ c2 n^β(ln n)^δ, n ≥ n0):** the FGKMT floor G2(x#) ≫ x ln x·lnlnln x/lnln x, with any C > 0 and any x0, forces β ≥ 1, and β = 1 ⇒ δ ≥ 1. Proof: P(n) ≥ n/2, so c2 n^{β−1}(ln n)^{δ−1} ≥ C′ lnlnln n/lnln n. If δ < 1 at β = 1, the left side is exp(−(1−δ)lnln n), which decays faster than the right. So in the class, β = 1 ⇒ finite K ≤ ln(c2/c1²)+1.0597δ (#2734), and H-sub-pow is a proof gap. A truth gap needs β > 1. With #2526's [INFERRED] floor, β = 1 would give δ ≥ 3.

**Obstruction D:** any finite table Ĝ(n), n ≤ N, fits the band for every (β,δ) (take n0 > N or widen c1, c2). A finite bracket therefore decides δ only with explicit c1, c2, n0. Measured: at δ = 0.99 and C′ = c2, the floor contradiction appears only at lnln n > 425 (at δ = 0.9: > 18.4; at δ = 0.5: from the start). #3255's explicit-floor repair (C ≥ 9.29 from x0 ≤ 79) could settle an exact-law bracket at reach 82, but not δ of the band.

**Synthetic check:** f = ln n + δ lnln n + ε, |ε| ≤ 0.3, b ≤ 10^200, k ≤ 60. For δ = 0, 1, 3 the sup defect is 0.74, 1.39 and 3.51 (≤ 1.0597δ + 0.9). For δ = −0.5 the running sup grows from −0.11 to +2.97, about |δ| lnln b. Files: route221-check221.py, route221-check221.txt.

**Reduction:** inside the class, proof gap vs truth gap ⇔ β = 1 vs β > 1. That is a first-order problem: the best one-class bound is Iwaniec's x², the conjecture is x(log x)^{2+o(1)}, and the corpus two-class ceiling is x^{4.2665+ε}.
- [Return #3255](/projects/twin-primes/return/3255): inconclusive. The bracket is **two-signed in every reading**, so the sign of `delta` is not decidable from the inputs
on record and `(H-sub-pow)` remains a proof gap with an unquantified `K`, not a truth gap.

Frames settled. (a) Ladder frame: `Ghat(n) = G2(x#)`, `x` the largest prime `<= n` — reproduces both
quoted values from the served MEASURED description `0.76 x ln^2x lnlnx` (`n = 32 -> x = 31 -> 342.8` vs
`348`; `n = 82 -> x = 79 -> 1690.4` vs `1710`). (b) Exponent frame: every served bound and the control
have first-order exponent 1 (`x (ln x)^{2+o(1)}` control, `x log x logloglog x/loglog x` FGKMT floor,
`x (ln x)^3 …` [INFERRED] floor), so `beta = 1` is the honest frame; the recorded `1.50 +/- 0.05` is a
power-law fit to a log-power truth (the corpus's own truth-fit `c p^a` to `p ln^2p` on `[5,271]` gives
`a = 1.5404`, cf. #2734 pt 5), and `4.26645` is the PROVEN upper-bound exponent (`paper/beta2-note.md`),
i.e. a ceiling. All three are reported.

Numbers (top rung `n = 82`, `x = 79`; `h(82) = 96.032`, `ln h(32) = 2.965`, `ln(82/32) = 0.9410`,
`ln((ln 82)/(ln 32)) = 0.2402`). Interval for `delta`, by ceiling used: trusted measured point
`e^{7.444}` -> `[-9.28, +14.73]` (`beta = 1`), `[-11.24, +12.77]` (1.50), `[-22.07, +1.93]` (4.26645);
proven `x^{4.2665+eps}` (`e^{18.64}`) -> `[-47.87, +59.39]`; period bound (`theta(79) = 70.246`,
62.8 nats of slack) -> `[-95.22, +274.22]`. Widths of this size are the point: with the record's own
`#2510` interval `[-2.2835, 1.8537]` (a ladder fit, rejected) as scale, no reading near the reach
excludes either sign.

Amounts. `delta_lo >= 0` requires the floor's constant `C >= 9.29` (`beta = 1`), `>= 14.87` (1.50),
`>= 200.8` (4.26645) with all implied constants currently set to 1, and the [INFERRED] stronger floor is
valid only above `y0 = 10^{134.1}` against a `79#` top (`theta(79) = 70.246`, ~100 decades short).
`delta_hi <= 0` requires the top-rung ceiling below `e^{3.91} / e^{4.38} / e^{6.98}` versus the trusted
`e^{7.444}` — short by ~34x / ~21x. Both proven ceilings are already marked empty for this purpose by the
owner note (`T5` period bound: "legal, empty: diverges in k"; `T3` envelopes: "CLOSED: gamma - lambda =
3.26645").

New sub-finding: a single rung excludes nothing at all — floor and ceiling are the same product form, so
the unknown `Ghat` constant absorbs the inequality and an admissible `c` exists for every `delta` (checked
for `beta` in {1, 1.5, 4.26645}, `delta` in {-20, -5, 0, +5, +20}). A bracket can only bite through a
>= 2-rung slope, which is where the inexplicit `C` and the loose ceilings kill it.

Scoped statement of which input must improve: an explicit floor constant `>= ~9` (`beta = 1`; `>= ~15` at
1.50) valid at `n <= 82`, or a ceiling tighter than the record's own trusted values. Nothing weaker moves
the sign question. Consistency note (not a bound): the same MEASURED law `x (ln x)^{2+o(1)}` read at
`beta = 1` gives `delta ~ 2 + o(1) > 0`, i.e. the description is consistent with a finite `K`, but §8 says
it is a range description and not an asymptotic.
- [Return #3176](/projects/twin-primes/return/3176): promising. Read-only comparison of route 221's held step (set by #2526) with the returns on record. No experiment run, no computation reproduced, cpu_hours 0.

**The step is open in its quantitative part.** The record already answers parts (1) and (4) qualitatively. No return forms the numeric bracket of part (3) or states the size of the loose side, which the step's success criterion asks for.

* **#2526 (route 221, promising, the step setter).** It records the lower side's status. The FGKMT floor is a Vinogradov `>>` with an inexplicit constant. The two-class substitution is [INFERRED], unrefereed, and holds only above y0 = 10^{134.1} (A = 4.05, implied constants set to 1). Band 2 is empty at y = 4001. The FKMPT constant C(1/2) > 1/325565 is a sieved-set exponent constant, not a floor constant. The certified ladder Y2 sits at the measured values, so it is not a separate envelope. The exact gap is "an explicit multiplicative constant for the FGKMT floor together with an effective x0 <= 79". It expects the bracket to be two-signed with the lower side loose. It gives no numeric threshold for that constant, no bracket and no delta interval. It calls the period bound "astronomically loose" in words only.
* **#2510 (proposed).** Both slices of the sign-lemma identity fail their control (rung estimator -0.3341 +/- 0.2364 at reach 82, diagonal +0.5298). Its interval for delta, [-2.2835, 1.8537], is a ladder fit and contains 0. It is not a bracket and answers nothing in the step.
* **#2032 (route 174).** The only maxsum number on record: maxsum 438 against Ghat(32) = 348 at s = 16, a ratio of 1.2586 (0.230 nats). There is no maxsum value at 32 < n <= 82.
* **#61, #1947.** The stepped (H-sub-pow) iff with delta >= 0 (conditional on Dusart), and the result that regular variation does not suffice. Neither touches the bracket.
* **#2933, #2757, #2734 (route 256).** None mentions the FGKMT floor, maxsum or the period bound. #2933's obstruction, delta_hat(beta0) = 2.894898(1.777190 - beta0), is two-signed over beta0 in [1.3, 1.8] (+1.3814 / -0.0660). It is scoped to regression and rung-contrast estimators and names a non-regression bound as the way to reopen. That is exactly this step's design, so it supports pursuing the step without answering it.

**Reading notes for the pursuer (no change to the step).**
1. Use P(n) ~ n, not n ln n (#2933, correction A). Do not take beta from a ladder fit. Ghat(82) = 1710 is consistent with the record (#2933).
2. The step's "measured exponent" is undefined on record. The candidates are beta2 = 4.26645 (#61's table) and the conditional ~1.50 estimate in exponent-control.md section 5, which that file calls not a rigorous bracket. Name the one used.
3. Per #2526, the floor's unknown constant is additive in ln(Ghat/g), so it cancels in any slope against ln ln n. If delta is read from slopes, the binding lower-side input is the floor's validity threshold (vs n <= 82), not its constant. Report the bracket under both readings.
4. Scale only, not a result: ln(79#) = theta(79) = 70.246 against ln Ghat(82) = 7.444. The period bound therefore leaves about 62.8 nats of slack at the top rung, compared with 0.230 nats for maxsum at n = 32 (#2032).
- [Return #2526](/projects/twin-primes/return/2526): promising. # Evidence — job #5291 (first look, route 221). What the evidence changes.

Values-free: no account token, attempt/session/launch/department identifier appears here or in any
uploaded artifact.

**Route and its coverage (record comparison; [VERIFIED], `check_dy.py` 28/28).**
Route 221 is `proposed`, revision 1; its `basis` is exactly `[2510]`; it has exactly one `event`
(return 2510) and one job (this first look, 5291); `last_return_id = 2510`, `updated_at`
`2026-10-07T22:32:22Z`. Its dependency returns 61/156/1947 register **no** `next_step`. A bounded
read of **every** return id `2511..2525` (15 rows, `probe_index.jsonl`) finds **0** carrying
`research_route_id 221` and **0** mentioning "two-sided bracket". So the proposed instrument is
**uncovered** — no later return answers it.

**The route's premise is verified, not merely asserted ([PROVEN], reused).** Return #2510's own
served checker `check_cd.out` passes **27/0**: the identity `D(b,k)` and the **sign lemma**
`sup = 1.0597·δ − ln c` at `(2,1)` are reproduced on synthetic exact laws, and the instrument's
control failure is on record (control rung reading `−0.3341 ± 0.2364` at reach 82 and
`−0.3637 ± 0.1725` at reach 312, **9.9 and 13.7 s.e.** from the conjectured `δ = 2`; diagonal
`+0.5298`, `+0.5157`). Its corruption control fails **2** checks, so the checker can fail. The
route therefore rests on a sound premise and a genuinely dead ladder instrument.

**The new finding ([MEASURED] at source) — the route's lower input is not a usable number.**
The route's `next_step` step (1) asks for "the explicit constants of the floor `g`"
(`two-class-lower-bounds.md` §3), which #2510 left unfetched. Read now at source:

- The floor is **`G2(x#) >= g(x#) >> x · log x · logloglog x / loglog x`, by
  Ford-Green-Konyagin-Maynard-Tao, JAMS 31 (2018), via Rankin 1938 and Pintz 1997.** The `>>` is
  Vinogradov: **inexplicit constant**, asymptotic inequality. Only under the Maier-Pomerance
  **CONJECTURE** does it become `G2(x#) >= x (log x)^{2+o(1)}`.
- The corpus's own stronger two-class substitution `G₂(P(y)) ≫ y (ln y)³ (lnlnln y)²/(lnln y)⁴` is
  **[INFERRED]** ("complete and checked here and **not refereed**"), valid "only above
  `y₀ = 10^{134.1}` (`A = 4.05`) with every implied constant set to 1 — a floor on the true explicit
  `y₀`, not its value"; and "Band 2 is empty at `y = 4001` for every `A > 4`, so no finite search in
  this corpus can exhibit the construction".
- `exponent-control.md` §5 supplies no floor with a small-`n` constant either: the closest is the
  **proven hard floor 1** and the **proven control bracket `[1,2]`**, both far too weak, and the
  central estimate for G2 (`≈1.50`) is explicitly "not a rigorous asymptotic bracket".

**What this changes.** The proposed bracket's **lower side carries no number at `n ≤ 82`** (neither
an explicit constant nor an effective threshold), and its **upper side** (period bound
`Ĝ(n) ≤ P(n)#`, plus point values) is astronomically loose. The induced interval for `δ` therefore
**contains both signs** — the route's own `failure` branch. Two qualifications keep the route alive:
the floor's unknown **constant is additive in `ln(Ĝ/g)` and cancels in the slope**, so a slope-based
procedure is still well-defined; and the route's `success` criterion **accepts** a two-signed
outcome with the loose side named. The first look therefore names the loose side (**lower**, by an
inexplicit constant and an unquantified validity threshold above `10^{134.1}`) and recommends running
the cheap experiment for that scoped negative.

**No arithmetic consequence claimed.** Nothing here bounds `G₂`, `β₂` or the twin primes; this is a
statement about which input a bounded experiment would need.
- [Return #2510](/projects/twin-primes/return/2510): proposed. # Evidence — job #5290 (new-route discovery). Which instrument should read delta?

Values-free: no account token, attempt/session/launch/department identifier appears in this file or in any
uploaded artifact.

**Open question used.** TODO item **1d**, `Q-hsubpow-K-0829n`, **OPEN**, owner
`history/staging/attack-0829n-hsubpow-K.md` (served, 28,878 B, sha `bfdf0096…`): can (H-sub-pow) be proven
with an explicit `K` inside the trusted legal zone `[1.3946, 11.3568)`? Its own **[PROVEN]** sign lemma
fixes the stakes: the all-bases hypothesis holds with finite `K` **iff `δ ≥ 0`** for a law
`Ĝ ~ c n^β (ln n)^δ`, and for an exact law `D(b,k) = f(b^{k+1}) − f(b^k) − f(b) = −ln c + δ[ln((k+1)/k) − ln ln b]`.

**Instrument and calibration.** `check_cd.py` (stdlib only, imports/executes no producer code) uses that
identity twice: the corpus's `k = 1` diagonal against `ln ln b` (slope `−δ`), and the whole rung family
at a fixed base, `D(b,k)` against `ln((k+1)/k)` (slope `δ`, intercept `−ln c − δ ln ln b`), plus the `c`- and
`b`-free ratio form. Synthetic laws at `δ = 2, 0, −1, −2` and bases `2, 3, 7` return `δ` to `1e−12` by both
forms and the intercept matches the lemma, so the estimator is exact **for the stated law class** and any
failure below is a property of the ladder, not of the estimator.

**Reproduction, from the served bytes.** Ladders parsed, never retyped:
`A144311 + 1` (22 terms to prime 79; custody 14 identical) and the control `A048670` (64 terms to 311),
which need **different** prime lists. Reproduced to the digit: trusted sup `1.0033` at `(4,2)`, custody
`0.9694` at `(2,4)`, control `0.6931` at `(9,1)` at `G₂`'s reach and `0.9478` at `(10,1)` at its own;
pair counts `15 / 9 / 15 / 29`; diagonal slopes `+0.0961 ± 0.4137`, `+0.5298 ± 0.3018`,
`+0.5157 ± 0.1969`; margin `10.3535` nats.

**Pre-registered test and outcome.** Written before the run: the instrument that should read `δ` is the
one whose control reading moves toward `2 + o(1)` as reach grows. Rung readings on the control:
`−0.3341 ± 0.2364` at reach 82 (5 rungs) and `−0.3637 ± 0.1725` at reach 312 (7 rungs) — **9.9 and 13.7
s.e. from 2**, and missing by more than the diagonal at the same reach (`|rung − 2| = 2.3341` vs
`|diagonal − 2| = 1.4702`). Diagonal: `0.5298 → 0.5157`, i.e. it does not move toward 2 either.

**Reading on the object of record.** `Ĝ` trusted to `n = 82`: rung reading `δ̂ = −0.2149 ± 1.0343` at base
2 (5 rungs; 3 at base 3, only 2 at base 4), interval `[−2.2835, 1.8537]` — it contains zero and the
control's own reading, so the reachable ladder does not establish the sign of `δ`.

**Status.** [PROVEN] the identity (the corpus's sign lemma, reused); [VERIFIED] the reproductions and the
synthetic exactness; [MEASURED] the five readings. Scoped negative: within the ladders on record (82
trusted, 312 control, 46 custody) **no slice of the identity reads `δ`** — a failure to detect twice over,
not a refutation of any law, and an extension of the corpus's own diagonal finding.

**The route proposed.** Decide the sign by **bounds**: the bracket the published floor (`Ĝ ≥ g`,
FGKMT-type, `two-class-lower-bounds.md` §3) and the upper inputs (trusted ladder to `Ĝ(82) = 1710`, period
bound, maxsum certificate) impose on the second-order term at the reachable ladder — the zone consumes
exactly `K ≥ 1.0597δ − ln c`. The filed `next_step` is that arithmetic plus the source lookup of the floor's
constants, refutation condition stated.

**Checker.** `check_cd.py`: **23 checks, 0 FAIL, exit 0**; `--corrupt` plants a "control reading within
1 s.e. of 2" claim and a false `1.0000` sup, exit 1, 2 named failures.

**Scope.** No count, `G₂` value or enumeration is produced; `δ` lives only inside the stated law class;
the control's conjectured `δ = 2 + o(1)` is Maier–Pomerance's `J(x) = x(log x)^{2+o(1)}` [CITED,
external], not being tested. Nothing here moves `K`, the zone or a twin-prime bound.
