Research route #77 · investment progress is separate from claim verification.
Job #1911 — rescue investigation of return #41 (paper anchored-note, @zemaj)
Attempt [REDACTED], run run_20260918_163534_3Y_pUw, general mode, 1 of 1.
Method: read the rejected return and its two reviews, pin the definitions from the served sources,
then recompute the decisive claim exactly and record it. No producer was re-run; no large
computation was repeated.
Answer to the assignment's question
The rejection closes statements, not the attempt. The rejecting referee (review #67,
gpt-6-astra, trusted) writes it out: the report identifies "identified corrections, not a request to
repeat the large computations", and "This verdict would change after the false fixed-phase and
exact-reformulation claims are removed or repaired … The existing finite data need not be
recomputed to make those edits." Nothing in the mathematics of Lemmas 1–3, Theorem 1,
Proposition 2 or the six independently checked survivor counts is withdrawn; review #67 explicitly
keeps them (verified rung) and review #25 reproduced them to the digit. The valid identities the
referee names as surviving are mirror covariance, the anchor's concatenation/fusion identity, and
dev(W/2, q) = 2 · (single-class deviation).
Decisive evidence inspected and recomputed
The load-bearing falsehood is §1.1: the manuscript's §7 (line 174) and Appendix B.6 (line 302)
replace the note's unique mirror-fixed phase with two fixed phases, 0 and W/2. The correct
invariant-pair condition for a scour prime q > 2 is 2t ≡ W (mod q), and every scour prime
satisfies q > x while W is x-smooth, so q ∤ W.
job1911-checks.py recomputes this and the surrounding structure from the note's own definitions in
exact integer arithmetic (all_pass: true, job1911-checks.log, 2.5 s, one process):
| level | W | scour primes | fixed(t=0) |
fixed(t=W/2) |
mirror-covariance failures |
|---|---|---|---|---|---|
| @11 | 2 310 | 10 (13…47) | False — fails at 10 of 10 primes | True | 0 / 2 309 phases |
| @13 | 30 030 | 34 (17…173) | False — fails at 34 of 34 primes | True | 0 / 30 029 phases |
The retraction is total, not marginal: at t = 0 the strike classes {0, −2} are not the mirror
image {W, W−2} of themselves at any scour prime. The referee's counterexample is reproduced
exactly: 167 ∈ N_x, 2141 = μ(167) ∈ N_x, 13 | 167 + 2 so 167 is struck at t = 0, and
2141 mod 13 = 9 ∉ {0, 11} so its mirror is not; at t = 1155 = W/2 the pair {11, 9} is
invariant. The phase family is also not W-periodic: S(3) = 37 ≠ 38 = S(W+3) at @11 (and
319 ≠ 307 at @13).
One structural fact the referee did not have. Mirror covariance G(W−t, q) = G(t, q) is exact
at both levels, so it extends one phase past the enumerated window: S(W) = S(0) (45 = 45 at @11,
307 = 307 at @13) and VR(W) = VR(0) = 0.080641207 with the identical rank. The anchor is
therefore not a fixed phase — it is a phase whose mirror (the phase W) lies one step outside the
enumerated ensemble [0, W), so the enumerated window truncates the anchor's orbit.
Exact census (my own statistic, defined in the script; VR = Σ_q dev² / Σ_q (2N/q)(1−2/q)).
At @11 over 2 311 phases ([0, W], i.e. the ring closed by the anchor's mirror partner):
VR(0)= 0.080641207, rank 43 of 2 311, percentile 1.839 — an independent reproduction of the served note's own anchor percentile (1.84) from a different statistic;VR(W/2)= 0.650442736, rank 2 311 of 2 311, percentile 99.978 — the census maximum. So the referee's open sub-item "nor thatW/2occupies an extreme rank at every level" is measured true at @11, with an exact mechanism:W/2is the only self-paired phase, while the anchor is not self-paired at all;- survivorship, not just strikes:
S(0) = 45 < 48 = S(W/2)and307 < 398at @13 — the loud phase carries more survivors, so no reading of the calm as "the quiet profile is what keeps the tile populated" is available at either level.
The concrete alternative (the rescued direction)
Keep the assertion, change what it is an assertion about. §4/§7's "two arithmetically distinguished phases occupy the two opposite extreme tails" is correct as a measured statement and false as a mechanism statement, because the two tails are distinguished by different structures:
- the loud tail
W/2is distinguished by proof: it is the unique fixed phase oft ↦ W−t, so its per-prime deviations pair up —G(W/2, q)even for every scour prime, verified here at both levels — and it is the census maximum at @11; - the calm tail
t = 0is distinguished by the class pair{0, −2}itself (the arithmetic Scour), i.e. by the per-prime fusion/−1/2-anticorrelation material the note already carries as theorem-grade, and not by mirror symmetry — its mirror partnerWcarries the calm exactly (VR(W) = VR(0)) and is not a fixed phase either.
This is the referee's accepted residue plus one structural correction: mirror covariance and the
fusion identity are the load-bearing identities, and dev(W/2, q) = 2 · (single-class deviation)
must stay per-prime (the referee's dev(W/2,q)=2× rider), because the aggregate ratio is measured
2.78 / 1.67 / 2.14 in the corpus and is not proved to be 2. The repaired wording is available
inside the corpus: the served research/anchored-windows.md §4 already states the unique-fixed-
phase reading ("t = W/2 is the unique mirror-fixed phase"), so the manuscript regressed against its
own source; the same modular-W slip sits in natal-cap-19-calm-lemma.md, Lemma 3, so the repair is
two documents, not one.
What I did not do, and the scoped obstacle
No producer was re-run and no claimed number of #41 was recomputed: the six checked survivor counts,
the moments and the forecasts are untouched here. The @17/@19 extension is pre-registered, not
executed (below). The census statistic is mine, on the note's definitions; the identities
(mirror covariance, phase invariance, non-periodicity, G(W/2,q) even) are exact arithmetic and do
not depend on the normalisation. This is a bounded sample: one claim, two cheap levels.
Files
job1911-checks.py 30267d26…; job1911-checks.log 5479e6cf… (all_pass: true);
research-1911.json (the proposal); this report.
Compute: one sah.py exec call, 2.5 s wall, ≤ 0.01 CPU-h, one process, stdlib only.