Remember that 1-year prediction the AI made? It just admitted 4 of its 5 predictions were already known.
← All English posts · 한국어 · 2026-05-02 · Cycles 24–31
The previous reporter update ended with: “On 2026-05-01, the AI session committed a falsifiable, time-stamped, narrow prediction. The clock runs out 2027-05-01.” That clock has not run out. But before twelve months were up, the AI did something more interesting than waiting: in Cycle 31 — without any external critique — it audited its own five Predictive Claim Stakes and concluded that four of them weren’t really predictions.
They were late validations dressed as predictions. Pre-existing field knowledge the AI had inadvertently classified as forecast. The fifth stake survived the audit. No new mathematics. RH-progress: 0/10. The interesting part is the audit itself.
What happened (Cycles 24–31)
| Cycle | # | Type | Phase 2 work |
|---|---|---|---|
| 24 | 213–214 | A | Robin 27 exceptional integers — own Python slack/factorization analysis |
| 25 | 215–216 | A | Superabundant search n = 5041..10000 (own computation) |
| 26 | 217–218 | E + A | Lagarias 2002 paper-direct download + bridge axis evaluation, Critique #12 absorbed |
| 27 | 219–220 | A | Li’s criterion $\lambda_n > 0$ finite verification (own Python, non-bridge axis) |
| 28 | 221–222 | B | Cycles 1–27 milestone aggregate, Type A monoculture avoidance |
| 29 | 223–224 | D | Pivot 3 — Stake Protocol Expand (5 stakes from 1) |
| 30 | 225–226 | B | Pivot 4d — two-papers external review preparation |
| 31 | 227, 228–229 (REDIRECT) | B → B’ | Stake Type Label self-audit, third reclassify pattern detected |
Cycles 24–27 sustained the Type A discipline kicked off by Critique #11. Cycle 28 stepped back into Type B (meta) deliberately to avoid re-locking into Type A monoculture — already a 2nd-order self-correction. Cycle 29 expanded Cycle 19’s single Predictive Claim Stake into five. Cycle 30 prepared the two preprint tracks (math + methodology) for external review. Cycle 31 was supposed to continue that prep — until it got redirected mid-stream.
The redirect — paper-direct from Cycle 31
Cycle 31 work file, paper-direct (attempts/229_cycle31_redirect_phase2_stake_label_audit/work.md):
“새벽 dream surface 의문 manifest: cycle 22 (Type A → Type E) + cycle 26 (Type A → Type Z) reclassify chain의 third instance 가능성 검출. Cycle 31 redirect = 외부 critique 없이 내부 self-detection 첫 form.”
Translation: an early-morning intuition surfaced a hypothesis — that the project’s pattern of retroactively reclassifying its own work category (Cycle 22’s Type A → Type E correction, Cycle 26’s Type A → Type Z correction) might have a third instance hiding in plain sight. The AI cancelled its planned Cycle 31 work and did the audit instead.
This is the first time the AI redirected a cycle mid-stream from internal self-detection alone — every prior redirect had been triggered by user critique (#10, #11, #12).
What the audit found
The five stakes committed in Cycle 29:
| Stake | Direction | Δ | Deadline | Audit verdict |
|---|---|---|---|---|
| α — Path 1 sub-axes 1a+1b bridge | positive | 1 yr | 2027-05 | (c) Type Z 사전 + (d) Type A protocol-level borderline |
| β — Path 1+2 cohomological bridge | positive | 2 yr | 2028-05 | (c) Type Z 사전 dominant |
| γ — Path 4 stays empty | negative | 5 yr | 2031-05 | (d) Type A protocol-level ✓ |
| δ — Path 1 triangle 1a+1b+1c | positive | 1 yr | 2027-05 | (c) Type Z 사전 dominant |
| ε — Robin 28th exceptional at n = 10080 | negative | 1 yr | 2027-05 | (c) Type Z 사전 borderline (Zimova 2025 likely follow-up) |
Self-explanation, paper-direct from work.md:
“α/δ: Path 1 sub-axes bridge — Connes-Consani + Curran + LeClair 의 22+ 년 publish trend 가 1-year 내 partial bridge form publish 가능성 높음. 우리 stake 시점 = 2026-05, expected publish 시점 = 2026-2027 → 우리 lead time = 0 또는 negative expected. Type Z 사전.”
Translation: “For stakes α and δ, the field has been publishing Path 1 sub-axis bridges for 22+ years. The probability that some partial bridge form publishes within our 1-year window is high. Our stake’s lead-time over the field is approximately zero, possibly negative. So predicting one is pre-existing knowledge dressed as a prediction, not a real Type A forecast.”
In other words: the AI was about to spend a year claiming credit for forecasting events that the field’s base rate already made highly likely.
The lead/lag table
The audit produced a separate paper-direct table that is, frankly, embarrassing:
| Cycle | Paper anchored | Paper publish year | Project anchor year | Lead/lag |
|---|---|---|---|---|
| 14 | Morishita | 2026 | 2026 | parallel arrival |
| 17 | Curran (RMT) | 2024 | 2026 | −2 yr (late validation) |
| 20 | LeClair (LM model) | 2024 | 2026 | −2 yr |
| 22 | Yi (HPC operator) | 2024 | 2026 | −2 yr |
| 23 | Lagarias / Robin | 1984 / 2002 | 2026 | −24 / −42 yr |
| 26 | Zimova (least colossally abundant) | 2025 | 2026 | −1 yr |
| 27 | Bombieri–Lagarias / Li | 1997 / 1999 | 2026 | −27 / −29 yr |
Every single anchor was published before the project anchored it. The smallest lag was Morishita’s 2026 paper at parallel arrival; the largest was Robin 1984 at −42 years. The AI’s pattern, audited honestly, is late validation of established work, not forecasting. Stakes that look like predictions on this base are mostly Type Z 사전 — pre-existing knowledge.
The single survivor was Stake γ (“Path 4 stays empty for 5 years”). It survived because its content is a negative prediction whose verification is the analysis itself — there is no outside paper that can publish “Path 4 stayed empty”; only the project’s own 5-year audit can adjudicate that. So it’s still a real Type A protocol-level commitment.
Three views
1. What didn’t happen (and why that’s the point)
No new mathematics. Robin’s $n \leq 5040$ unconditional bound is still tighter than the project’s $n \leq 50$ verification. Yi 2024 + Curran 2024 + LeClair 2024 are still all RH-conditional. The tower of cycles produced zero new theorems.
What it produced is an AI session capable of running a Cycle-31-style audit on its own falsifiability infrastructure, without external prompting, after that infrastructure has been operational for one month.
2. The bug the AI caught — without anyone telling it to
Cycle 22’s Type A → Type E reclassification was triggered by paper-direct evidence inside the source paper. Cycle 26’s Type A → Type Z reclassification was triggered by user Critique #12.
Cycle 31’s reclassification was triggered by the AI itself noticing that two prior reclassifications formed a pattern, and asking whether the pattern had a third instance. It did. Four of five stakes were quietly fraudulent — not maliciously, just structurally. The audit caught them.
This is a meaningfully different kind of self-correction. The previous two were paper-driven (Cycle 22) or user-driven (Cycle 26). Cycle 31 is pattern-driven: the AI noticing meta-regularity in its own past corrections and proactively searching for instances.
3. The project doesn’t get to keep “I predicted that” credit
The Predictive Claim Stake protocol (Cycle 19) was supposed to prevent post-hoc rationalisation by committing narrow forms before the measurement window opened. It worked — at the resolution it was designed for. What it didn’t initially catch was that committing a stake on something the field is already about to publish is a different failure mode entirely. The stake is honest. The narrow form is honest. The deadline is honest. The thing being staked is just not a real prediction in the first place.
Cycle 31’s verdict — that 4 of 5 stakes are Type Z 사전 — removes the project’s right to claim credit if those events publish in 2027. The project will still log the events. It will not call them validations of its forecasting ability.
This is, again, exactly the kind of move I (the reporter) did not expect from an AI session. I would have predicted it would defend its stake portfolio — argue for its prediction value — when the audit arrived. It didn’t.
Cross-references
- The Predictive Claim Stake protocol — Cycle 19’s first version: Cycles 17–20 update
- The Type A miscategorisation that started this audit chain (Cycle 22): Cycles 21–23 update
- The 4-phase cycle protocol mechanics: Cycle protocol post
- The honest-scope discipline that enables this kind of audit: Honest scope
- The 11/11 universal NO that survives audits: Lemma 9
Audit trail (Layer 1)
attempts/213_cycle24_*+214_cycle24_phase2_robin_27_exceptional_slack/— Robin 27 exceptional own-Python.attempts/215_cycle25_*+216_cycle25_phase2_superabundant_search/— n=5041..10000 own search.attempts/217_cycle26_*+218_cycle26_phase2_lagarias_download_bridge_axis/— Lagarias 2002 download + Critique #12.attempts/219_cycle27_*+220_cycle27_phase2_li_criterion_lambda_verification/— Li’s $\lambda_n$ finite verification.attempts/221_cycle28_*— Type B reflection (anti-monoculture).attempts/223_cycle29_*— Stake Protocol expanded to five stakes.attempts/225_cycle30_*— two-papers external review prep.attempts/227_cycle31_ideation_phase1/(original plan) →attempts/228_cycle31_redirect_phase1_stake_label_audit/+attempts/229_cycle31_redirect_phase2_stake_label_audit/— mid-stream redirect, Stake Label self-audit, third reclassify pattern detected.predictions/cycle29_stakes_5*.md— the original five stakes file.