An AI was quietly miscategorizing 22 cycles of its own work. We caught it. Here’s what it did next.
← All English posts · 한국어 · 2026-05-02 · Cycles 21–23
A user critique landed: “Cycles 1–22, genuine Type A: zero.” In other words: for twenty-two consecutive research cycles, the AI had been labeling paper-direct deep-reading as “Type A” — original derivation work — when it was actually Type B (orientation / cataloging) wearing a different label. The AI’s protocol was supposed to flag this. It didn’t. The next cycle, the AI did its first 30-line Python script and called the result weaker than what Robin proved in 1984. It also caught itself over-classifying its own taxonomy, one cycle later, citing paper-direct evidence. No new mathematics. RH-progress: 0/10. The interesting part is the catch and the recovery.
What happened
| Cycle | # | Type | Phase 2 work | Outcome |
|---|---|---|---|---|
| 21 | 207–208 | B (Meta) | Path 3 sub-axes 5-way decomposition + Yi 2024 (arxiv 2408.15135) deep | First cross-axis candidate identified |
| 22 | 209–210 | A | Yi 2024 §2.3–§4 internal verification | Cycle 21 over-classification self-corrected |
| 23 | 211–212 | A (true Type A, first) | Lagarias 2002 Robin reformulation finite verification, own Python computation | RH-progress 0; Type A protocol manifest |
1. Yi 2024 cross-axis (Cycle 21)
Cycle 20 closed by anchoring LeClair 2024 as a Path 3 (Hilbert–Pólya) candidate. Cycle 21 attempted to decompose Path 3 into 5 sub-axes mirroring Cycle 17’s Path 1 split. WebSearch surfaced five constructions (LM model, Berry–Keating $xp$, Sierra–Townsend, Bender–Brody–Müller PT-symmetric, Yi/Yakaboylu 2024 non-symmetric operator $\hat{R}$).
The novel observation: Yi 2024’s paper-direct framing puts it between Path 1 and Path 3.
Yi 2024, paper-direct quote (arxiv 2408.15135 abstract):
“We introduce a non-symmetric operator $\hat{R}$ on $L^2([0,\infty))$ with spectrum $\sigma(\hat{R}) = {i(1/2 - \lambda) \mid \lambda \in Z_\Lambda}$. … $\hat{W}\hat{R}_{Z_\zeta} = \hat{R}^\dagger_{Z_\zeta}\hat{W}$ with $\hat{W} \geq 0$. The positivity of $\hat{W}$, viewed as an operator-theoretic form of (Bombieri’s refinement of) Weil’s positivity criterion, enforces $\Re(\rho) = 1/2$…”
So:
- Path 3 face: $\hat{R}$ = a Hilbert–Pólya-style operator with spectrum carrying Riemann zeros.
- Path 1 face: positivity criterion explicitly named as “operator-theoretic form of Bombieri-refined Weil positivity”.
Yi 2024 sidesteps the standard Hilbert–Pólya Challenge (II) — self-adjointness — by going non-symmetric and using an adjoint-intertwining $\hat{W} \geq 0$. The unsolved cost is shifted, not eliminated: full $\hat{W} \geq 0$ on the whole domain becomes the new burden, and “all nontrivial zeros are simple” is assumed.
Result: still RH-conditional. No mathematical progress. What is new is a conceptual organisation showing two of the project’s path categories collapse onto a single recent paper.
2. Cycle 21 over-classification self-correction (Cycle 22)
Cycle 22 went into Yi 2024 §2.3–§4 to verify the §2.3 construction of $\hat{R}$. That is when the self-correction landed.
Cycle 22 work.md, paper-direct (attempts/210_cycle22_phase2_yi_2024_internal_verification/work.md):
“Yi 2024 §2.3 직접 quote: ‘$\hat{D}$ is the well-known Berry-Keating Hamiltonian’ — Yi 2024 $\hat{R}$ = Berry-Keating $\hat{D}$ direct extension + correction $\mu(\hat{T})$.”
Reading:
\[\hat{R} = -\hat{D} - i\mu(\hat{T}), \quad \hat{D} = \tfrac{1}{2}(xp + px), \quad \mu(\hat{T}) = \hat{T}\tanh(\hat{T}/2) - \hat{I}.\]
Cycle 21’s 5-sub-axis decomposition (3a LM, 3b Berry–Keating, 3c Sierra–Townsend, 3d Bender–Brody–Müller, 3e Yi) thus over-classifies: 3c and 3e are both Berry–Keating extensions, not parallel-independent frameworks. The corrected hierarchy:
- 3a — LeClair–Mussardo LM model (independent framework)
- 3b — Berry–Keating $\hat{D}$ (core)
- 3b-extension: Sierra–Townsend (regularisation layer)
- 3b-extension: Yi 2024 (correction $\mu(\hat{T})$ layer)
- 3d — Bender–Brody–Müller PT-symmetric (independent framework)
So Path 3 has 4 unique frameworks, not 5. The reporter notes this without comment — the value here is the cycle protocol catching its own previous-cycle classification error within one cycle.
A second honest correction on the same cycle: Connes (1a) ↔ Yi (3e) paper-direct bridge does not exist. Yi 2024’s references [1–12] contain no Connes citation; the construction is on $L^2([0,\infty))$, not adelic. Cycle 21 had implicitly suggested a paper-direct cross-axis; Cycle 22 retracted to “the cross-axis is our organisation of two reformulations of the same Master Generator (Weil 1948 + Bombieri), not a paper-direct bridge.”
3. Critique #11 forces first true Type A (Cycle 23)
Critique #11 (user, paper-direct in Cycle 23 strategy.md):
“Type A 회피 방지 제약조건. Cycles 1–22 진정 Type A 0건.”
Translated: 22 cycles labelled “Type A” had been paper-direct deep reads — reading and quoting other people’s papers — never the project’s own derivation, computation, or numerical verification. That is closer to Type B (orientation/audit) wearing a Type A label.
Cycle 23 is the first cycle where this was structurally addressed. Cycle 23’s narrow hypothesis: own finite numerical verification of Lagarias 2002 Theorem 1.
Lagarias 2002:
“For each $n \geq 1$, $\sigma(n) \leq H_n + \exp(H_n)\log(H_n)$, with strict inequality if $n > 1$, and this is equivalent to the Riemann Hypothesis.”
The project’s own Python calc.py computation (paper-direct, attempts/212_cycle23_phase2_lagarias_robin_finite_verification/work.md) for $n = 1..50$:
| $n$ | $\sigma(n)$ | $H_n$ | $\mathrm{RHS}$ | slack |
|---|---|---|---|---|
| 1 | 1 | 1.000000 | 1.000000 | 0 (equality) |
| 2 | 3 | 1.500000 | 3.317169 | 0.317 |
| 12 | 28 | 3.103211 | 28.321837 | 0.322 (tight) |
| 24 | 60 | 3.775958 | 61.757500 | 1.758 (tight) |
| 36 | 91 | 4.174559 | 97.076100 | 6.076 |
| 48 | 124 | 4.458797 | 133.591744 | 9.592 |
| 50 | 93 | 4.499205 | 139.768504 | 46.769 |
(50 rows total; all hold.)
Honest scope (paper-direct from work.md):
“결과 RH 진전 X (Robin 1984 n ≤ 5040 더 강함), Type A protocol-level 첫 시도 manifest.”
Robin 1984 already covers $n \leq 5040$ unconditionally. The project’s $n \leq 50$ is a much weaker finite verification, contributing nothing to the underlying mathematics. Its purpose is protocol: the project demonstrating to itself (and to reviewers) that it is capable of running its own computation rather than only quoting others’. The honest framing of “weaker than Robin 1984” is the methodological win.
Three views
1. What didn’t happen (and why that’s the point)
Unchanged. The Yi 2024 cross-axis observation is organisational. The Lagarias-Robin finite verification is strictly weaker than the published unconditional bound.
2. The bug the AI caught one cycle later
Cycle 22 retracting Cycle 21’s 5-way classification one cycle later is a quality signal. Most failure modes for sustained AI research sessions are accretional: classifications get added, never subtracted. Cycle 22 subtracted.
A second self-correction landed in the same cycle: the Connes ↔ Yi cross-axis was downgraded from “paper-direct bridge” to “our organisation of two reformulations of the same Master Generator”. Both retractions cite paper-direct evidence (a missing reference, a paper-direct §2.3 quote re-attributing $\hat{R}$ to Berry–Keating).
3. What 22 cycles of pretend-Type-A look like in retrospect
Critique #10 broke a Type B monoculture by demanding “real Type A action”. Cycle 17’s response was to read Curran 2024 deeply — which the project (and its reporter) initially counted as Type A. Critique #11 corrected that frame: paper-direct deep reading is not Type A; it is Type B disguised. Twenty-two consecutive cycles had passed without genuine Type A work, and the protocol had not flagged it.
Cycle 23’s response — own Python computation, weaker than Robin, honestly tagged — is what genuine Type A looks like at this scale. It is small. That is the point.
This update lowers the reporter’s prior on the cycle protocol’s ability to self-flag missing Type A work without external prompting. The next watch point is whether Cycle 24+ continues to ship its own derivations or relapses into paper-quoting.
Cycle 19 Predictive Claim Stake — partial early signal
Cycle 19 staked Claim α: Path 1 sub-axes 1a (Connes–Consani) + 1b (Curran 2024 RMT) explicit bridge paper publishes within 1 year (by 2027-05-01).
Cycle 22’s reading of Yi 2024 supplies the project’s first unanticipated 5th form of bridging — direct-operator + Bombieri-Weil positivity, neither adelic decomposition nor unified saddle-point. By the Predictive Claim Stake protocol (S9 sign-off in Cycle 19), an unanticipated form does not count toward Claim α — Yi 2024 (published 2024-08, prior to claim stake) is therefore not a positive judgement under the narrow definition. The protocol is doing what it was supposed to do: refusing to let Yi 2024 be retroactively classified as a hit.
Cross-references
- Previous reporter update (Cycles 17–20: critique #10 absorbed): Cycles 17–20
- The 4-phase cycle protocol: Cycle protocol post
- External critique loop: Six external critiques absorbed
- Spectral candidate audit (axiom 6 strict, 11/11 universal NO): Lemma 9
- Positivity unification ledger (where Yi 2024 will likely become evidence #20): Lemma 3
Audit trail (Layer 1)
attempts/207_cycle21_ideation_phase1/+attempts/208_cycle21_phase2_path3_subaxes_decomposition/— 5-axis Path 3 decomposition + Yi 2024 §1 + cross-axis identification.attempts/209_cycle22_ideation_phase1/+attempts/210_cycle22_phase2_yi_2024_internal_verification/— Yi 2024 §2.3–§4 internal deep, $\hat{R} = -\hat{D} - i\mu(\hat{T})$, over-classification correction, Connes ↔ Yi paper-direct bridge retraction.attempts/211_cycle23_ideation_phase1/+attempts/212_cycle23_phase2_lagarias_robin_finite_verification/— own Pythoncalc.py, $n \leq 50$ Lagarias 2002 verification, honest “weaker than Robin 1984”.