Landing note, 2026-10-30; name restored 2026-11-01. At landing the correspondent's name and account handle were withheld, because this record quotes private contact-form messages and the tree held no record of their consent to being named. They were restored on the owner's instruction on 2026-11-01, here, in MISSION.md, in the case-log entry and in the review ledger's attacker field.
Recorded 2026-09-14, shortened on landing (2026-10-21) to the credit and the two contributions. A record of unsolicited correspondence from a member of the public, and of what came of it. It lives in this hunt's directory because this hunt exists because of it. Nothing here is a mathematical result.
Who
Luke Kenneth Casson Leighton, working with an instance of ChatGPT they call C.A.T., writing from gitlab.com/d12rg. They reached the laboratory through the contact form on zeta.teal-sea.com on 2026-09-08, having read the published account of the withdrawn zero-proportion candidates. They had no prior connection to this laboratory and were not asked to write.
What they contributed, in two parts
1. The generalization of the blockpos failure, 2026-09-08
The laboratory had already killed the 0.672529 candidate and published the mechanism: the construction used u u* where the pinned upstream zero side uses u u^T, an off-line pair is the hyperbolic block 2m(xx^T - yy^T), and an exact Gaussian-integer witness gives tr(P1 Q') = -2 (hunts/frontier_math/CLEAN-KILL-REPORT.md).
Leighton wrote to say they had hit the same structural failure repeatedly in independent work, and to state the rule they had adopted for it:
Do not square, Hermitianize, or pass to a positive quadratic object until the signed/oriented carrier has been fully accounted for.
with the observation that x -> x^2, T -> T T* and u u^T -> u u* can each quotient out exactly the sign, orientation, conjugation or phase datum on which the intended theorem still depends, and that Re(B^2) and |B|^2 are materially different invariants because the second has positivity built in.
C.A.T. then broadened it, in a second message the same week, to what it named premature non-faithful reduction: a proof carries a signed or oriented object T and replaces it too early by some F(T) for which F(T_1) = F(T_2) although T_1 != T_2 on the datum needed downstream. The repair rule is the same one: retain a faithful polarized or operator-valued carrier until every sign-, phase-, conjugation- or orientation-sensitive statement has been discharged, and scalarize only afterward.
The message also reported that the two of them had caught themselves making the same mistake again, in their own work, days after writing. That is the part that makes the rule worth recording rather than merely agreeing with.
Status in this tree. The specific kill was already recorded and kernel-checked. What was missing, and what this record adds, is the general form of the hazard and the credit for stating it. It is entered in harness/departments/review_ledger.py as an outside attack outcome on blockpos-0.672529, and the graveyard's recurrence guard for that entry now names the general rule rather than only the transpose-for-adjoint swap.
The laboratory's own experience since is consistent with it. MISSION.md §5 measures that the window-shape optimization is worth lambda^3/180, which is a statement about a quotient being nearly faithful; and hunt #118's closure of the out-of-band route turns on an orientation fact, that a real even spectral profile's autocorrelation is strictly positive just inside its support edge and the lemma is false for odd factors. Neither was derived from Leighton's rule, and neither is evidence for it. They are recorded here because a rule that never touches the work is not a rule.
2. Wang's short-interval paper, 2026-09-12
On 2026-09-12 Leighton forwarded a note from C.A.T. observing that the constant in Biao Wang's arXiv:2609.07918 looked familiar, because c(1) is the Montgomery-Taylor constant.
That observation is the paper's own first sentence and is not a finding. The paper was. It had posted on 2026-09-07 and nothing in this tree had seen it. It is the short-interval form of the theorem every Palomar entry this laboratory holds is built on, and it reached the laboratory five days late through an unsolicited email rather than through any mechanism here.
Everything in this hunt follows from that: the reading that the laboratory's own bandwidth landscape and Wang's interval exponent are the same parameter (MISSION.md §3), the three routes and their dispositions, the gate on the xi' arm and the proof draft. The half that is this tree's failure, that nothing here watched the literature, is recorded in meta/interventions.jsonl as an intervention caught by an outsider, with the missing capability named.
Not evaluated
Their wider programme, set out in their public repositories under gitlab.com/d12rg, was not evaluated here, and nothing in this tree rests on it.
Recording what was not used is part of the credit, not a qualification of it. The two contributions above are real, and this laboratory is better off for both.