Tools chosen: polymarket_direct claude_news gdelt_news claude_news wikipedia polymarket_related kalshi_related code_execution

Sub-questions (Fermi decomposition)

Planner reasoning

This is a Polymarket question about whether Epoch AI officially marks at least one additional FrontierMath: Open Problems problem as 'Solved' before Sept 30, 2026. The key drivers are the base rate of prior solve announcements (how many of the ~untouched open problems have been solved and how fast), whether Epoch is actively adding problems/promoting the program, and AI capability trajectory (GPT-5.x/Gemini-class models, Erdos-problem-style solve waves). The market price is the primary anchor, supplemented by news on Epoch AI announcements.

# Current state Polymarket's mirror of this exact market has surged to **90.95% YES**, up ~41 points in the last 7 days (identical to its 30-day change, implying the entire price history is only ~6-7 days old). This surge lines up with two Epoch AI "Open Problems" classification updates in the second week of August 2026 (a human-credited solve on 8/11 and an AI-credited provisional solve on 8/12) — either of which, if it post-dates market creation, would already satisfy this market's YES condition. The remaining ~9% gap to 100% likely reflects residual uncertainty over whether Epoch's provisional classification will hold, and/or ambiguity about exact market-creation timing relative to these events. # Timeline of key events - 2024-11: Epoch AI establishes FrontierMath benchmark (confirmed, Wikipedia). - 2026-01-27: Epoch AI launches "FrontierMath: Open Problems" pilot, 14 unsolved research problems (reported, claude_news). - 2026-03: First Open Problems solve — a Ramsey-hypergraph problem, elicited from GPT-5.4 Pro by Kevin Barreto/Liam Price, based on Will Brian's unsolved 2019 paper (confirmed, Epoch substack). - 2026-06-12: Epoch corrects errors in 42% of FrontierMath problems, rescoring models (reported, Digg/GDELT). - 2026-07-31: Set expanded from 14→50 problems; the March Ramsey-hypergraph solve is **removed** for insufficient notability, leaving "3 AI solves, 6%" cited as of this date (confirmed, Epoch/cryptobriefing). - 2026-08-11: Inverse Galois problem (Mathieu Group M23) marked **solved by humans** (Huang et al.) — counts under Epoch's classification but not as an "AI solve" (confirmed, Epoch page). - 2026-08-12: Hadamard matrix of order 668 marked **solved (provisional)**, credited to 3 humans + Claude; Epoch flags it may be revised (confirmed, Epoch page). - 2026-08-13: "Book Ramsey Numbers" problem — Dualverse AI candidate construction under author review, **not yet confirmed solved** (reported). - ~2026-08 (exact date unclear): Polymarket market created; price history begins around 50%, now 90.95%. # Event Will Epoch AI mark at least one additional FrontierMath: Open Problems problem as "Solved" (or officially state one is solved) between market creation and Sept 30, 2026? # Outcomes to forecast - Yes - No # Kalshi market anchor No Kalshi-direct price was returned for this ticker; kalshi_related search found no matching FrontierMath/AI-math market. **Primary cross-market anchor is Polymarket: 90.95% YES**, up from a 50% low, with only $15.4K volume over ~6 days — thin but directionally strong and recent. # Sub-question answers 1. **Set size/solves to date** — Open Problems expanded from 14 (Jan 2026) to 50 (Jul 31, 2026); as of ~Aug 12, 2026 there have been 3 AI-credited solves (one later delisted) plus one human-credited solve (M23, 8/11) and one provisional AI solve (Hadamard-668, 8/12) (claude_news, cryptobriefing, Epoch.ai). 2. **Historical solve rate** — Roughly 1 credible solve every ~2-4 months from March–August 2026 (λ≈0.25–0.5/mo). Poisson modeling (code_execution) gives P(≥1 solve) ≈20-63% over a 1-2 month remaining window depending on λ. 3. **Post-creation announcements** — Yes: Epoch's page shows classification updates on 2026-08-11 (M23, human) and 2026-08-12 (Hadamard-668, AI, provisional), plus an active near-solve (Book Ramsey Numbers) under review as of 8/13 — timing relative to exact market creation is the key unresolved question. 4. **Active expansion/bounty** — Yes; Epoch expanded the set 14→50 problems (7/31/26) and continues correcting/adding problems; Schmidt Sciences funds the Open Problems track independently of OpenAI's Tier 1-4 funding (cryptobriefing). 5. **Broader AI-on-math trajectory** — Strong and accelerating: GPT-5.2 Pro solved Erdős #397 and #281 (Tao-verified), OpenAI claimed an 80-year-old Erdős unit-distance conjecture solve (May 2026), DeepMind's AlphaProof Nexus solved 9 Erdős problems cheaply, and Tier 4 FrontierMath solve rate rose from 3 problems (mid-2025) to 20/48 (42%) by March 2026 — all indicating rich adjacent solving capacity, though Tao cautions these are "lowest-hanging fruit." 6. **Polymarket/related markets** — This exact market on Polymarket: 90.95% YES (see anchor above). Manifold priors (pre-dating this window): "new Open Problem solved before Aug 2026" priced ~30%; "≥2 solved before 2027" priced ~82% — consistent with steady-but-not-fast cadence. # Key facts (high-confidence, factual) 1. [Epoch.ai] Open Problems set = 50 problems as of 2026-07-31. 2. [Epoch.ai/claude_news] 8/11/26: M23 inverse Galois problem marked solved (human-credited). 3. [Epoch.ai/claude_news] 8/12/26: Hadamard matrix order-668 marked solved (AI-credited, provisional). 4. [Wikipedia] First Open Problems solve was a hypergraph problem by GPT-5.4 (later delisted from benchmark for notability, not overturned as unsolved). 5. [cryptobriefing] "AI solved 3 of ~50" was the state as of the 7/31 expansion, before the two August events. # Cross-market signals - Kalshi related: no direct FrontierMath market found; one unrelated SCOTUS market surfaced (irrelevant). - Polymarket: 90.95% YES, up from 50% low, thin volume (~$15K), 6-day history — strongly bullish post-August-events repricing. - Manifold (proxy, pre-window): ~30% for "new solve before Aug 2026"; 82% for "≥2 solves before 2027" — implies market participants see solves as fairly regular but not guaranteed monthly. # Analyst opinions and speculation - Tao (via multiple outlets) cautions AI math "solves" are often literature retrieval or lowest-hanging fruit, and attribution (human vs AI) is contentious — Epoch's own M23/Hadamard classifications show this ambiguity directly. - Claude_news synthesis argues cadence (roughly monthly across FrontierMath + Erdős-adjacent work) makes another solve "quite likely" by Sept 30, 2026. # Directional lean per outcome - **Yes**: Supported by (a) Polymarket at 91%, (b) two Epoch classification events (8/11, 8/12) very close to/likely after market creation, (c) accelerating AI math capability trend, (d) active expansion of problem set increasing surface area, (e) an additional near-solve (Book Ramsey) under review. - **No**: Opposing — provisional Hadamard classification could be reversed; short remaining window (~1-2 months) relative to historical λ implies base-rate Poisson probability of only 20-60%; genuinely novel unattributed AI solves remain rare (only a handful all year). # Gaps / unknowns - Exact market creation date/time not confirmed — critical for whether 8/11 and 8/12 events already count as "additional" solves under this market's rules. - No Kalshi-direct price available for calibration; relying solely on Polymarket (thin volume, short history). - Whether the Hadamard-668 provisional solve will be finalized or revised (Epoch explicitly flagged reversal risk). - Status of the Book Ramsey Numbers candidate solve (unresolved as of 8/13). # Calibration anchors - Polymarket YES 90.95% (current, thin volume) — primary anchor in absence of Kalshi-direct data. - Poisson base-rate model: λ≈0.25-0.5 solves/month → 20-63% over remaining window, notably below Polymarket price, suggesting either market has already priced in near-certain recent trigger events, or is overconfident. - Manifold historical analog: 30% for "new solve before Aug 2026" (pre-expansion baseline).

1. Both forecasters anchor almost entirely on Polymarket's 90.95%, but the brief notes this is a thin market ($15.4K volume, ~6-day history) that could be driven by a handful of informed or uninformed traders—neither forecaster discounts for potential manipulation or small-sample noise despite flagging it as "thin." 2. Neither forecaster grapples with the critical asymmetry in the "March Ramsey-hypergraph solve" precedent: it was *removed* from the set for insufficient notability months after being credited, showing Epoch's classifications are reversible even after being publicly counted — this directly undercuts confidence that the 8/11 and 8/12 events will "count" toward this specific market's resolution criteria, yet both treat them as near-certain triggers. 3. Both forecasts effectively assume the 8/11/8/12 events occurred *after* market creation without strong evidence — the brief explicitly states "exact market creation date/time not confirmed" and the Polymarket price history starting near 50% could equally reflect a market created *after* awareness of these events, meaning the "already happened" thesis may be double-counting information already priced in rather than a genuine edge over the crowd. 4. The Poisson base rate (20-63%) is presented as the "bearish counterweight" but neither forecaster explains why they weight it so lightly (roughly 10-20% weight) versus the market signal — given historical solve cadence is genuinely the best forward-looking data point (excluding the ambiguous August events), a more rigorous blend might land meaningfully lower than 87-93%. 5. Neither forecaster addresses resolution-criteria ambiguity: if the market requires an "additional" solve strictly after creation and the Hadamard-668 provisional solve is later reversed (which Epoch explicitly flagged as possible), this could flip a currently-assumed YES trigger to a No outcome with the remaining window then relying solely on the weaker Book Ramsey candidate or fresh solves — this reversal risk is understated in both rationales beyond a passing mention. 6. Both converge tightly (0.87-0.93) despite acknowledging major unresolved uncertainties (creation timestamp, provisional status, thin volume), which suggests anchoring bias toward Polymarket rather than independently derived confidence intervals — true uncertainty given the gaps listed should arguably produce more spread or lower confidence scores than 0.61-0.66.

89% (64% confidence)

Yes: 89% No: 11%

The Polymarket mirror at ~91% with a 41-point surge coinciding exactly with Epoch's 8/11 (M23, human-credited) and 8/12 (Hadamard-668, AI-credited provisional) classification updates strongly implies the triggering event has already occurred post-creation, which dominates the pure Poisson base rate (20-63%) that ignores realized events. The critique's strongest points — reversibility of Epoch classifications (the March Ramsey-hypergraph solve was delisted), the explicitly provisional Hadamard status, and unconfirmed market-creation timing — justify shading modestly below the anchor rather than a large discount, since even under a full discount of the August events there remains a ~1.5-month window with an active Book Ramsey candidate under author review, a newly expanded 50-problem surface area, and accelerating AI-math cadence. Thin volume ($15K, 6 days) argues for humility, but the price direction is corroborated by independently documented Epoch page updates, not just trader sentiment. I settle at 0.89 Yes, just below the consensus midpoint and the market anchor, reflecting genuine but bounded resolution-criteria and reversal risk.

Total pipeline time: 240.0s