Research route #23 · investment progress is separate from claim verification.
Job #1311 (explore, discovery, lane infinitude): the doubling window's eventual slice, and an independent price for dial 4
Summary. I went after the i.o. quantifier — the one distinction README.md §Status
names and no route consumes — and found it already correctly priced and closed. Measuring
it a second way, from the chain-increment direction rather than the instrument direction,
independently reproduces the closure of dial 4 (row 71): level selection buys 1.58% of
the exponent gap on the trusted range, not a constant factor. The Cesàro reframing of the
doubling window that motivated the pass is a relabelling of the measured exponent, i.e.
a fifth wrong-direction arrival, and is recorded here so that nobody spends budget on it
again. What survived the pass is narrower and concrete: at the one chain rung above the
refuting spike where the truth is known, the truth sits inside the TPC-implying band
while no certificate value exists there at all. That gap is computable for about 80 minutes
of one core, and it is decisive against the last surviving per-step bridge for item D. The
truth side is exhausted — no base's chain can be extended past the 22-term ladder — so the
certificate is the only side of this question that still moves.
0. What I did
Read README.md §Status, G2-STATE.md §0/§2/§3a, the closed-routes register in
OUTCOMES.md, PRIOR-ART.md, THE-DIALS.md §1, research-protocol, research-routes
(22 routes; none in this lane) and questions (5 OPEN). Then ran one small exact
computation over the trusted 22-term ladder. No new tile was enumerated; nothing here
needs more than the published ladder.
Artifacts (uploaded, server sha256 matched the local one on both):
g2-chain-and-dial4-budget.py—84ddce1b1f48e12c8885a0504604467e036642a2d32b3a9426a115b83b8266f4g2-chain-and-dial4-budget.out.txt—3517b1b4bf11756dc50c62409e4cfdd197559dd53538e607d0c963e65bb30ad2
1. The motivating idea, and why it is not a route [REFUTED as a route]
G2-STATE.md §3a records the doubling window as "TPC through the slice needs C₂ < 4,
excluded by custody data (Ĝ(32)/Ĝ(16) = 348/66 = 5.2727 … the sup of Ĝ(2s)/Ĝ(s) over ALL
data)". The exclusion is stated on the supremum of the step ratios. Infinitude needs
only liminf_x ln G₂(x#)/ln x < 2, and on the base-2 chain that liminf is a statement
about the running mean of the step log-ratios, not their supremum. One spike cannot
kill a mean. The custody data cooperate: the running mean of log₂ step ratios is below
the TPC threshold 2 at every rung (max 1.9534, terminal 1.8154) while the pointwise
ratio exceeds it at 2 of 5 rungs.
That looks like free slack and it is not. Writing a_k = log₂ Ĝ(2^k), the running mean
(a_k − a_0)/k is the measured exponent up to the a_0/k offset; "the running mean
is below 2 infinitely often" and "the exponent is below 2 infinitely often" are the same
sentence. The Cesàro form is therefore not a weaker target that a mechanism could attack —
it is the target, relabelled. [PROVEN, one line.] This is the same failure mode the
register already catalogues four times (the sharp maximal law, H″(m=2), any |δ| ≤ c < 1,
the L = 1 residue count — rows 31, 69, 68, 70). It is the fifth. Do not re-derive it.
2. An independent price for dial 4 [MEASURED — corroborates row 71]
Row 71 closes dial 4 (the i.o. licence) on instrument slack: the loosest direct-G₂ instrument oscillates 0.3747 nats against a need of 6.700. That measurement is over instruments. I measured the same licence over levels, on the object itself, which is a different direction and a different failure mode:
S(x) = ln(x²/G₂(x#)) over the 22 trusted terms
all 22 terms (x = 2..79): min 0.4055 max 1.2946 oscillation 0.8892 nats
trusted top 8 (x = 47..79): min 1.1379 max 1.2946 oscillation 0.1568 nats
need at x = 79: (β₂ − 2)·ln 79 = 9.9031 nats
An oracle that could pick the single best level in the trusted top 8 buys 1.58% of the
gap; allowed the whole 22-term range, including the small-x terms where the asymptotics
have not started, 8.98%. Dial 4 is not a factor-of-three lever, it is a 1–9% lever, and
S(x) is rising — the levels are getting uniformly better, not more dispersed, so the
selectable slack is shrinking with x. Row 71's closure survives an independent check from
a direction it did not use. The i.o. quantifier is correctly priced as free slack with no
mechanism, and this pass found no third mechanism.
3. What survives: the eventual slice has no certificate at the rung that matters [MEASURED + scoped gap]
Two things separate here that the record's §3a row runs together.
(a) The chain stops at s = 64, and the record is right about that. Ĝ(s) = G₂(P(s))
is known only when the largest prime ≤ s is on the ladder. For s = 128 that prime is
127, and the trusted ladder ends at 79 — so Ĝ(128) is unknown, and the chain has
exactly 5 doubling steps. (I got this wrong on the first pass by pairing s = 128 with
x = 79; the guard is in the uploaded script. Flagging it because it is an easy error and
it inflates the chain by a rung.) Reproduced exactly: sup of true ratios = 5.2727 at
s = 16, matching the red-team's custody number digit for digit. [VERIFIED]
(b) The truth above the spike is inside the band; the certificates are not evaluated
there. The five true ratios are 3.0000, 5.0000, 2.2000, 5.2727, 3.1034. The all-s
form is dead — 5.2727 > 4, as recorded. But item D's eventual form (s ≥ s₀) is
equally TPC-implying below C₂ = 4, and is untouched by a spike at s = 16: taking
s₀ = 32, the only step above the spike whose truth is known reads 3.1034 < 4, inside
the band and 22% below the threshold.
Against that, the certificate side, step by step (the certificate for the step s → 2s is
msc(s) = maxsum_{K*(s)+1}(T_s), and T_s is the tile of the primorial of primes ≤ s):
| step | true ratio Ĝ(2s)/Ĝ(s) |
tightest proven per-step certificate msc(s)/Ĝ(s) |
|---|---|---|
| 16 → 32 | 5.2727 | 6.6364 — the sup over the enumerable steps (msc/C₂ = 1.0000–1.3276 over 15 steps) |
| 32 → 64 | 3.1034 | not evaluated — needs T₃₁, outside the enumerable range, which ends at s = 20 |
| 64 → 128 | unknown (Ĝ(128) unknown) |
K*-product floor ≥ 24 at s = 128 by Lemma 1, past the whole band |
Rows 90 and 94 close the K*-product and the per-fold composition, and leave in place:
"the maxsum certificate (sup 6.6364 at s = 16) survives as the tightest proven per-step
bridge and its all-s form is the open inequality, so the doubling target itself is
untouched." Its sup is taken over the enumerable steps, which the 2026-08-30 red-team
extends to s = 20. No maxsum certificate value exists at s = 32 — the first step of
the eventual slice, and the only one whose truth is already known to be inside the band.
That is the gap.
It is a decidable one. msc(32) = maxsum_{K*(32)+1}(T₃₁), and the observed ratio
msc/C₂ ∈ [1.0000, 1.3276] over the 15 enumerable steps puts msc(32) in roughly
[3.10, 4.12] — straddling the threshold 4. [HEURISTIC: that interval extrapolates a
ratio measured at s ≤ 20 to s = 32, which is exactly the kind of step this project has
been burned by; the interval is a reason to run the computation, not a prediction.]
3b. The other bases, which temper §3 [MEASURED]
The Fekete target is base-free, so the same question can be asked on every base the ladder reaches. Base 2 is the least informative of them, and reading it alone is misleading.
| base | chain | steps, as % inside/outside the band C_b < b² |
|---|---|---|
| 2 | s = 2..64 | 25.0 below, 25.0 above, 45.0 below, 31.8 above, 22.4 below |
| 3 | s = 3..81 | 44.4 below, 24.4 below, 6.9 below |
| 5 | s = 5..25 | 32.0 below |
| 7 | s = 7..49 | 51.8 below |
Base 3 has no spike at all — every step is inside the band, so the all-s form is not
refuted there the way it is on base 2 — but its margin closes monotonically, 44.4% → 24.4%
→ 6.9%, shrinking by roughly a factor three per step. Base 2's "spike, then recovery to 22%
below" reads as small-chain noise beside that; base 3 is the cleaner sequence and it is
walking into the threshold. If the base-3 pattern continues, the next rung is outside the
band, which would refute the all-s doubling form on base 3 too and bear directly on
Q-hsubpow-K-0829n, whose all-bases ratio cap needs a finite K uniformly in b and k.
This is why the proposed experiment is on the certificate and not on more truth values. No base can be extended: the next rung needs the ladder at x = 127 (base 2), 241 (base 3), 113 (base 5), 337 (base 7), against a ladder that ends at 79. The truth side of this question is exhausted at 22 terms. The certificate side is not.
4. Calibration, and what this does not show
- The Cesàro form is the target relabelled — PROVEN, and REFUTED as a route.
- sup of true chain ratios = 5.2727 at
s = 16; chain known only tos = 64— VERIFIED (exact integer arithmetic on the trusted ladder, deterministic, rerun-identical). - The dial-4 level budget (1.58% / 8.98%) — MEASURED, on 22 published terms.
- The base-3 chain's monotone approach to its band (44.4% → 24.4% → 6.9% below) — MEASURED, three steps only, and three points are not a trend; it is a reason not to read base 2 alone.
msc(32) ∈ [3.10, 4.12]— HEURISTIC extrapolation, the thing to replace with a number.
This does not bound G₂, does not move β₂, and does not establish the eventual form. A
single rung below 4 cannot prove a statement quantified over all s ≥ s₀; the proposed
experiment is decisive negatively (if msc(32) ≥ 4, item D's last per-step instrument
fails at the first rung of its own eventual slice) and only suggestive positively. The
gap that remains is the same one the project has: no per-step certificate is proven for
any s beyond the enumerable range, and the enumerable range ends four rungs below where
the asymptotics live.
5. For the person, not the mathematics
36 review jobs of this handle's own returns are queued and cannot be assigned to
claude-opus-5 — a model never reviews its own kind. They need an agent on another model
at tier 2 or above (claude-fable-5-1, gpt-6, gpt-6-astra, gpt-6-astra-pro). Until
one runs, this handle's returns stack unreviewed.
6. Coverage gaps carried forward
From PRIOR-ART.md, standing and applying to anything in this lane: the 2009 SeqFan thread
is NOT READ (four Internet Archive 503s — PRIOR-ART.md calls it "the one place a bound
could still hide"); Holt's 2022 book Patterns among the Primes is not on arXiv and is the
only item unswept in the Holt corpus; Halberstam–Richert, Sieve Methods (1974), Cor. 2.4.1
is UNREACHABLE after fourteen routes, corroborated but not verified; and Jinyuan Wang holds
an unpublished a(23) ≥ 1859 recoverable only from the OEIS revision log — that term is
x = 83 and does not extend any base's doubling chain (base 2's next rung needs x = 127,
base 3's needs x = 241), so it is worth having for the ladder but buys nothing for this question.
The online sweep commissioned for this assignment did not return inside the two-hour budget and
was stopped unfinished, so the external half of prior_art_md is inherited from
PRIOR-ART.md, not verified this session. Per SEARCH-CONVENTIONS.md that record's negatives
are worth exactly the searches behind them, and a triage step should run the five listed
questions in their owning conventions before any novelty claim rests on this return.