Harvest record — the twin-prime residue-patterns chat

Filed as a record so the harvest survives, not only its outputs. Kept verbatim below the line; this block is the outcome.

Outcome, 2026-09-14. §1 filed as probes/P-008-no-exceptional-scale-dyadic-chowla.md, type lead (the discriminating test is a proof, not a computation). §2 filed as probes/P-009-two-bets-one-statement.md with related: CASE-07, P-005 and relation: undetermined — whether it is the same mechanism is the probe's own work, not a call made at filing. §3 filed as notes/log-weight-haar-measure.md — the positive statement the take's false-friends log recorded only as a "No". §4 filed in OPS.md (inductive) as the third instance of converge your control, the first from outside the dynamics work. §5 (the stated preference) recorded to memory, not a repo registry. §6 amended into lab/runs/026-liouville-scale-up.md: its prediction and falsifier, moving it from draft to open. Note: the digest's own filing summary named P-006/P-007 for §1/§2; those ids were taken between the chat (2026-09-12) and this harvest, so the probes are P-008/P-009.

---

Harvest — "Twin prime residue patterns mod 4"

Source: chat 90bc0a56-3620-4e8c-8b62-87d605ca9b2e, 2026-09-12. Artifacts already in the repo: takes/twin-prime-dial.md, liouville_1e9.py, scripts/twin-dial/ (29 scripts).

What crossed and what didn't. The take page captured the exhibits, the false-friends log and the scripts. What stayed in the chat is what the templates had no field for at the time: the probes, one method lesson, and a correction to a shared picture.

Confirm each before filing — a retroactive harvest can't distinguish your decisions from Claude's suggestions as reliably as a live one.

---

1 · Probe — the missing lemma, stated precisely

It is written out in the chat, not paraphrased:

> There is no scale Y for which (1/Y)Σ_{Y<n≤2Y} λ(n)λ(n+h) ≫ 1 while the same quantity > is o(1) on average over scales within log-distance (log Y)^{1/2} of Y.

Structurally parallel to "no exceptional character." Data at 10⁷: dyadic values all at 1/√Y, block variance 1.00, no scale stands out.

Why it is a probe and not a case: the discriminating test is a proof, not a computation. The numerics have already said what they can — they cannot distinguish "no exceptional scale" from "no exceptional scale below 10⁷".

2 · Probe — the two bets are halves of one statement

Helfgott–Radziwiłł gives expansion at every scale. MRT uniformity at every scale would forbid the exceptional scales. So the two live bets are one statement seen twice.

The argument in the chat: c(y) is Lipschitz in log y with constant ≈ 2, so the only thing compatible with every window average is small and c(y) ↛ 0 is a sparse set of exceptional scales of log-width ≍ 1, where λ(n+1) ≈ λ(n) across a whole dyadic range. At such a scale λ has locally constant runs — which is a large short-interval Fourier coefficient, i.e. a failure of MRT uniformity at that scale.

Whether that implication can be made rigorous is the open question. Tao–Teräväinen's derivation of Chowla from uniformity still lands on the log-average.

Cross-reference, relation undetermined: two obstructions turning out to be one statement is the same shape as CASE-07 and P-005. Whether it is the same mechanism is what the work would settle, not a call to make at filing.

3 · Correction — where the logarithmic weight enters

Not the entropy step. The first identity.

> (1/p)Σ λ(pn)λ(pn+p)/(pn) = Σ_{p|n} λ(n)λ(n+p)/n needs the weight 1/n to make dilation exact > — it is the Haar measure of the scaling flow.

Everything after it is pointwise in scale: the expansion theorem holds for every N, and MRT's input holds for every N. That changes what the log-average is an obstruction to — it is not a technical artifact of a proof step that might be removed, it is the invariant measure of the group the whole picture is built on.

Recorded in the take's false-friends section as "the log-average comes from the entropy step — No." Worth having separately because the positive statement is the useful half.

4 · Method lesson — inductive, and a second instance of an existing check

Block series must be first-differenced before correlating. A headline result died to it: the apparent signal was a density trend, not structure.

That is converge your control before you trust your contrast in a different costume — the uncorrected series is a control that was never converged. The check currently stands at two instances (run 024's truncation, run 022's offset spread). This is a third, and the first from outside the dynamics work, which matters for a check about how errors arrive in this project.

5 · Stated preference

> Stop when the numerics have reached their limit, rather than generating further > confirmations.

Both leverage routes — the Heath-Brown Siegel-zero mechanism and the chi-twisted sieve — were tested and closed numerically, and the thread stopped. It did not go looking for a fourth confirmation of λ's randomness.

6 · Runs already scoped in the chat, for run 026's fields

The chat names what 026 is for, which the run file currently lacks:

• settle the triple-correlation bias Σλ(n)λ(n+2)λ(n+4) at 10⁹ — at 10⁷ the small-p excess is not a residue-class effect (chi-square is flat) but a global √y-scale bias
• per-scale information to p ≤ 10⁵ and Fourier uniformity to H = (log x)⁴

And the 10⁷ baseline is the falsifier's other half: λ matched a coin to three decimals from H = 8 through H = 65536, spanning the MRT threshold (log x)^{1+ε}. A departure at 10⁹ would be the first structure found; agreement extends direct evidence for the MRT conjecture at its conjectured threshold.

---

Filing summary