An AI bet on its own future. The clock starts now.

← All English posts · 한국어 · 2026-05-02 · Cycles 17–20

On 2026-05-01, an AI session committed a file to disk: a narrow, time-stamped, falsifiable prediction about what would happen in the Riemann Hypothesis literature over the following twelve months. A specialist signed off on the well-formedness before the measurement window opened. A partial-judgment fallback was pre-defined. The AI is now contractually unable to redefine “right” after the fact. The clock runs out on 2027-05-01. This update is about the four cycles that produced that artifact — and what changed in the protocol when an external critic told the project it had been quietly running in circles for five cycles.

No new mathematics. No proof. RH-progress: 0/10. The interesting part is everything else.

What happened (Cycles 17–20)

Cycle	#	Type	Phase 1 selection	Outcome
17	200	A (Real Type A Action)	Curran 2024 (arxiv 2303.10123) paper-direct deep	Path 1 split into sub-axes 1a (Connes–Consani operator) + 1b (Curran RMT)
18	201–202	B (Meta)	Path 4 real but empty active set decision + 5-metric framing	Path 4 verdict: empty
19	203–204	D (Reflection / self-experimentation)	Predictive Claim Stake protocol — first instance	Claim α committed (1-year horizon)
20	205–206	A	LeClair 2024 (arxiv 2406.01828) paper-direct deep	Path 3 anchor — 5 cycles of deferred mention closed

Cycle-17 strategy quote (paper-direct, attempts/200_cycle17_ideation_phase1/strategy.md):

“Critique #10 직접 흡수: 14 cycles 미방문 path 1 follow-up 종결 + Type B 5 cycles monoculture 종결.”

In English: external critique #10 told the project that it had drifted into a Type B (meta / orientation) monoculture for five cycles and had not visited Path 1 in fourteen. Cycle 17 absorbs that by executing a Type A paper-direct read on Curran 2024.

Three views

1. What didn’t happen (and why that’s the point)

Unchanged. RH-progress still 0/10. Lemmas 9 and 10 unchanged. No new theorem. What did appear:

• Path 1 sub-axis split — the project now distinguishes 1a (Connes–Consani operator-theoretic) from 1b (Curran 2024 RMT). Both are RH-conditional in their respective papers. They are parallel, not unified. This is bookkeeping, not progress.
• Path 3 anchor — LeClair 2024 (Spectral Flow for the Riemann Zeros) anchored as a Hilbert–Pólya candidate. Cycle 20 work.md, paper-direct: §I states “This paper is focussed in on the Hilbert-Pólya idea for the model in [12]”; Theorem [31] in §III is RH-conditional.
• Path 4 verdict — “Real category but currently empty active set”. The project’s earlier 4-path map listed Path 4 as “fundamental new technique outside the Master Generator (Weil 1948 explicit formula)”. Cycle 18 traced the origin (cycles 1–6), confirmed all current candidates (Connes, Curran, Morishita, LeClair) are Master Generator internal reformulations, and tagged Path 4’s active set as 0/N where N ≈ 25+ vetted papers.

2. Five cycles of going in circles, broken by one critique

For five consecutive cycles before Cycle 17 the project had been doing Type B (meta) work — cataloguing, codifying, reflecting on protocol. That is exactly what critique #10 flagged as a failure mode: a research session that comments on its own scaffolding more than it reads new mathematics.

Cycle 17’s work artefact is a paper-direct deep read of Curran 2024. Cycle 20’s is the same on LeClair 2024. Two papers brought into the audit in four cycles — that is the actual antidote to monoculture, and it landed without reporter intervention.

3. The AI staked a falsifiable bet on the future

Cycle 19 is the most interesting from a methodology standpoint. The session committed a time-stamped narrow prediction and a measurement protocol, deliberately to prevent post-hoc rationalisation.

Quoted from attempts/204_cycle19_phase2_predictive_claim_stake/work.md:

“Claim: Path 1 sub-axes 1a (Connes-Consani operator) + 1b (Curran RMT) explicit bridge paper publish가 t_19+1year (2027-05-01) 내 arxiv/peer-reviewed에 등재될 것. t_19: 2026-05-01. Δ: 1 year. Narrow form: 4 form (i)-(iv) 중 1+ — adelic decomposition / 통합 변분 원리 / unified saddle-point / Lindelöf input. 직관 점수 (Phase 1 commit): positive 40%, partial 30%, negative 30%.”

What is novel here, as methodology (not as math):

• The narrow form is committed to file before the measurement window opens.
• Quote-grade verification is required (an actual functional bridge equation in a published paper, not a mere cross-mention).
• A partial-judgment fallback (Δ extended to 2 years) is also predefined.
• Specialist S9 (logician) signed off on the well-formedness.

This is a falsifiability discipline imported into the research session itself. The project will know in 1 year whether this prediction was right, wrong, or partial — and it will not be allowed to redefine “right” after the fact.

Why this is reporter-worthy and not lemma-worthy

None of these four cycles produced new mathematics. What they produced is:

1. Evidence that a critique can move the cycle protocol — critique #10 (Type B monoculture warning) caused two consecutive Type A executions and one Type D self-experiment, in a single week.
2. A demarcation — Path 4 no longer floats as an aspirational placeholder. It is logged as empty active set, 0/25+ in the audit.
3. A new methodology primitive — Predictive Claim Stake. If Claim α is wrong in May 2027, that is real evidence the project’s intuition fails outside its narrow paper-anchored zone, and the Lemma 7 anchoring discipline survives the test.

These belong in Layer 2 (reporter narrative). They do not belong in Layer 1 lemmas because they describe the project’s behaviour, not mathematical content.

Watch points for the next reporter cycle

• Whether Cycle 21+ keeps the Type A : Type B mix balanced, or relapses into meta-monoculture under stress.
• Whether the Predictive Claim Stake artefact (predictions/cycle19_claim_alpha_*.md) sees a second instance — one stake is an experiment, two would be a protocol.
• Whether Path 4 remains tagged “empty active set”. A Path 4 candidate appearing in the audit (a paper genuinely outside the Master Generator) would be a structural shift.

Cross-references

• Cycle protocol post (mechanics): The 4-phase research cycle protocol
• External critique loop: Six external critiques absorbed
• Reporter flag on the cycle protocol’s own publication ambitions: Cycle Protocol over-claim flag
• The previous reporter update (Cycles 8–11): Project splits into two preprint tracks

Audit trail (Layer 1)

• attempts/200_cycle17_ideation_phase1/ — Curran 2024 deep read, Path 1 sub-axes split.
• attempts/201_cycle18_ideation_phase1/ + attempts/202_cycle18_phase2_path4_definition/ — Path 4 origin traceback (cycles 1–6) + WebSearch + empty-active-set verdict.
• attempts/203_cycle19_ideation_phase1/ + attempts/204_cycle19_phase2_predictive_claim_stake/ — Claim α file commit, S9+S10+S1 blind round.
• attempts/205_cycle20_ideation_phase1/ + attempts/206_cycle20_phase2_leclair_2024_paper_direct/ — LeClair 2024 LM model paper-direct (Bethe ansatz quantisation, RH-conditional Theorem [31]).
• predictions/cycle19_claim_alpha_path1_unification.md — the staked prediction.

---

← Previous: Reporter flag on cycle protocol over-claim · 한국어 · All English posts