The 722 Manuscript Problem: An Auditor's Read of OpenAI's Mathematical Claims

On October 6, 2026, OpenAI disclosed a result from an unnamed reasoning model. Input: 4,000 unsolved problems from the open mathematics literature. Output: 722 manuscripts. Clustered: 372 distinct result families. Average inference time per result: approximately three hours. Some outputs carry machine-checkable proof certificates in the Lean theorem prover. Many do not. I have been auditing machines for eight years. When a system produces a claim of value at scale, I do not read the press release. I trace the ledger. The ledger here is thin: one tweet, a single crypto media outlet relaying it, one unnamed advisor offering "an exciting day" as independent commentary. No model identity. No parameter count. No Lean coverage ratio. No peer review. No replication. No independent mathematician has confirmed a single result. This is not a breakthrough. This is a claim under audit. The distinction is professional, not semantic. Here is what the numbers say when treated as evidence rather than as marketing. The announcement belongs to a lineage. I audited my first smart contract in 2017 with the 0x Protocol v2 codebase, three months of static analysis that surfaced an integer overflow in the order-matching engine and delayed a launch by six weeks. I learned then that rigor costs speed and saves value. OpenAI's o1 series introduced extended chain-of-thought reasoning. DeepSeek-R1 demonstrated that reinforcement learning with verifiable rewards could train a model to reason at nearly frontier level for a fraction of the training budget. Google DeepMind's AlphaProof reached IMO gold-medal-level performance, with Lean as the verification backbone. In early 2024, I audited a DeFi protocol that proposed to let AI agents automate yield farming. The oracle layer lacked cryptographic verification of the AI's input data. Yield manipulation was possible. The project pivoted to a hybrid model with zero-knowledge proofs for data integrity. That experience became my template for assessing AI claims: trace the data flow, find the verification layer, determine whether the output can be checked. The 722 manuscripts sit where three industrial trends intersect. First, test-time compute scaling: the shift from larger models to longer inference runs. Second, formal verification infrastructure: Lean and other proof assistants moving from academic tools to industrial machinery. Third, the evaluation frontier: GSM8K, MATH, and AIME have been saturated, forcing frontier labs to search for harder scoring grounds. OpenAI's announcement is the first public, quantified snapshot of that intersection. It says nothing about the model. It says everything about the pipeline: massive parallel generation, plausible-output filtering, semantic clustering. That pipeline is the product. The crypto ecosystem should care. The same stack — long-duration inference, machine-checkable proof systems, high-throughput GPU clusters — underpins the next generation of AI finance products. The same language model that produces a candidate theorem produces a candidate trade hypothesis. Both are unverified generative outputs. Both require a verification layer before liability attaches. The announcement's pivot toward investment research — the closing movement of the coverage — is not an editorial accident. It is the commercial destination. Over the last year, the industry's marginal infrastructure dollar has shifted from training to inference. Long-duration reasoning queries are the new demand driver. 2,166 equivalent serial inference hours is a deliberate signal to that market. The math is the marketing. The compute is the product. Here is the systematic teardown. The pipeline mathematics deserve the first look. Read the input-output structure forensically. 4,000 problems in. 722 manuscripts out. 372 result families. The ratio 722 divided by 372 equals 1.94. That number is the most revealing figure in the entire disclosure. A system producing independent solutions would show a one-to-one relationship between manuscripts and results. 1.94 manuscripts per family is the fingerprint of massive parallel sampling followed by semantic deduplication: false starts, partial derivations, auxiliary lemmas, competing routes to the same mathematical destination. The model is not discovering like a mathematician. It is sampling from a distribution of candidate derivations and clustering whatever survives a lightweight plausibility filter. That distinction is categorical. A mathematician who solves a problem produces one proof. An automated sampler produces a manifold of candidate arguments. The sorting layer determines the quality. "Result families" are clusters of outputs, not certified theorems. The problem-level yield is 18.05%: 722 manuscripts from 4,000 problems. That is an upper bound on pipeline throughput before any correctness filtering. "Manuscript" does not mean "solution." It means "an artifact that looks like a contribution." Some of these artifacts are partial progress. Some are intermediate lemmas. Some are confident errors. The distribution is undisclosed. I have seen this shape before. During the Terra/Luna collapse investigation, I cross-referenced Anchor Protocol's 19% APY against its on-chain emission schedule. The yield was not generated by trading fees; it was newly minted LUNA distributed to depositors. It took fifty pages of transaction logs to demonstrate what the whitepaper had dressed in different clothing. Ponzi schemes leave trails in the data. So do unverified AI outputs. The trail here leads from the 4,000-problem claim to the 722-manuscript claim, and stops exactly where the verification ledger should begin. The verification gap is the core finding. "Many of the manuscripts include rigorous proofs in Lean. Others do not." In audit language, "many" is a disclosure of non-quantified partial coverage. An engineer who says "many functions have unit tests" is not reporting a coverage ratio. A security researcher who says "many endpoints are protected" is not telling you which ones are exposed. The choice of "many" over a percentage is the disclosure's most honest flaw. Lean is a proof assistant. Its kernel checks each step of a formal proof individually. A Lean certificate is the strongest form of mathematical guarantee currently available anywhere in computational science. It is not a claim about a proof; it is a machine-executable check. It is also expensive to produce. Formalization forces a strict grammar on the model, and converting a natural-language argument into a formal one frequently takes longer than the derivation itself. Given that cost asymmetry, the rational disclosure behavior for a laboratory with high Lean coverage is to publish the number. The absence of that number implies the coverage is not high enough to publish. My prior: fewer than half of the 722 manuscripts carry complete Lean certificates. The remainder are mathematical hypotheses rendered in confident natural language. They are indistinguishable in appearance from proven results until a human or a formal kernel checks them. The dangerous category is not the obviously wrong manuscript. It is the subtly wrong manuscript that looks plausible. Unverified AI output of that shape is the epistemic equivalent of an unverified smart contract that appears to implement ERC-20 correctly. The appearance of correctness is precisely what makes the failure expensive. "Output is not truth," the reporting notes. That sentence is the most important finding in the entire story. It should be printed on every AI product license. Audit the edges, not just the center. The center of this announcement is the 372 families. The edges are the unverified manuscripts leaking into preprint servers, academic citations, and investment research desks. The compute ledger carries its own accounting. Three hours per result. 722 manuscripts. Equivalent serial inference: 2,166 hours. Roughly ninety days of continuous single-GPU operation. At cluster scale, wall-clock time collapses; compute cost does not. The economics deserve scrutiny. Long-duration inference requires H100-class hardware, extensive KV-cache management, continuous batching, and fault-tolerant scheduling. The per-run cost of a three-hour inference is substantial. OpenAI disclosed the duration and withheld the aggregate. That asymmetry is intentional. Three hours is a capability signal. The dollar figure is a business constraint. Publishing one without the other is a selective disclosure. The generation-to-verification cost ratio is the structural fact. Generating a natural-language proof draft is compute-expensive. Checking a formal proof in Lean is cheap by orders of magnitude. In a fully automated pipeline, generation dominates the cost. But OpenAI's pipeline is not fully automated. The "others do not" category lands on human verifiers. Human verification is the most expensive input in the entire system. This asymmetry defines the economics of AI mathematics for the next decade. Generation scales with GPU procurement. Verification scales with human expertise and formal tooling. The verification burden is the true cost center. The lab — or the industry — that industrializes verification at the lowest cost per certified result owns the margin. I have watched this exact dynamic play out in crypto. The cost of generating a yield narrative has collapsed to near zero. The cost of auditing the underlying contract remains high. The premium accrues to the auditor. Audit fees for meaningful protocols remain stable in bear markets for exactly that reason. The same logic applies here. Complexity is often a disguise for theft. In this case, the complexity is the pipeline; the theft is the substitution of output counts for certified results. The epistemic ledger is where the systemic risk actually lives. This is not traditional AI safety. There are no jailbreaks, no rogue agents, no autonomy failures. The risk is epistemic: the production of plausible-looking mathematical claims at a speed that exceeds the ecosystem's capacity to verify them. 722 manuscripts. Unknown Lean coverage. The unverified remainder will enter preprint servers. They will be cited. They will anchor subsequent work. Some will be wrong. That is not a prediction; it is a statistical inevitability. The mechanism is not novel. In crypto, we watched unaudited contracts carry billions in total value locked because the user interface was convincing. We watched a centralized exchange display customer balances that did not exist on its ledger. The failure mode is identical: the surface presents confidence, the verification layer is absent, and the market prices confidence rather than verification. OpenAI's own warning — that some results "could have issues" — is simultaneously a risk disclosure and a liability shield. A lab that has publicly warned of possible errors can cite the warning when errors surface. The acknowledgment is accurate. It should also be read as structural: the warning exists because the verification gap is material. The correct operational posture for downstream consumers of AI mathematical output is the posture I recommend to institutional clients evaluating unaudited DeFi protocols: assume compromise until proven otherwise. Not because the actor is malicious. Because the absence of a verification layer means undetected error is materially probable, and in high-stakes environments, the rational prior is distrust. Silence is the only honest ledger. The announcement's silence on the verification rate is the most honest thing it contains. The industry aftershocks will be felt before the math is verified. The most durable effect is the shift from a generation economy to a verification economy. Verification becomes a distinct industry vertical: formal verification toolchains, Lean ecosystem development, automated proof checking, professional human verification services. The 722 manuscripts will matter less than the tooling that emerges in response to the bottleneck. Three effects are near-term. First, research infrastructure. The "AI research orchestrator" role — AI generates candidate hypotheses, humans prioritize, formal or semi-formal verification filters — becomes a standard workflow in quantitative research and investment analysis. The orchestration layer is the interface between generative machinery and the proof burden. Second, investment research. Models capable of sustained multi-hour reasoning over corporate filings, macro data, and competitive scenarios will restructure the buy-side research desk. The throughput gain will be real. The error rate will be non-zero. The firms that win will be those that install verification between model output and position sizing. The framing of an AI testing more hypotheses than any single analyst could is accurate about throughput. It is silent on the validation layer. That silence will be expensive for early adopters. Third, cloud demand shifts from training to inference. Long-duration reasoning workloads create a distinct demand profile: fewer, longer jobs, higher memory bandwidth, persistent KV-cache state. That shift is structural, not cyclical. The financial sector's adoption curve will be constrained by a boundary the coverage implies but does not specify. Mathematics is a closed system. Proofs are mechanically checkable. Markets are open systems. There is no Lean kernel for "the market will rise if the Fed cuts." The same model that produces a verified theorem cannot produce a verified prediction. AI's comparative advantage is in closed, verifiable domains. In open, nonstationary, adversarial domains — markets, politics, negotiation — the advantage shrinks and the error rate becomes exploitable. I have tested this boundary directly. I have run language models against on-chain forensics. They can summarize a transaction trace. They cannot audit one. The summaries sound authoritative. The audits require a verification layer that does not exist inside the model. The difference is the same difference between the 722 manuscripts and the Lean-certified subset. Output is not truth. The market is the largest open system in existence. Treat every AI-generated investment thesis as an unverified manuscript until independently checked. The competitive board is being redrawn in real time. OpenAI's statement that existing math benchmarks have become less useful is a public admission of benchmark saturation. GSM8K, MATH, AIME — all effectively solved. The standard evaluation suites no longer discriminate between frontier models. The field is responding by inventing evaluation frontiers: unsolved problems and open conjectures serve as the new scoring ground. Choosing "number of open problems converted to manuscripts" as the metric is a specific competitive strategy. It favors a lab with scale. It penalizes a lab that publishes smaller, verified output sets. Google DeepMind's AlphaProof line routes through Lean end-to-end. Its published results carry formalization as a matter of discipline. A verify-weighted comparison of the two pipelines would rank the labs differently than the raw manuscript count does. I therefore treat the choice of disclosure metric as an admission. A lab confident in its verification rate publishes verification rates. A lab that prefers raw output counts is optimizing the denominator it controls. Verify the hash, trust no one. The competitive dynamic has shifted from reasoning benchmarks to verification infrastructure. Lean is no longer an academic tool; it is competitive infrastructure. The laboratories that dominate the next phase will be the ones that build the highest-coverage verification pipeline, not the ones that generate the most manuscripts. The battle is for the certification layer. The also-rans matter less than the structural conclusion. Frontier labs are no longer competing on which model can answer a test question. They compete on which can convert generative capacity into certified knowledge with the lowest contamination rate. The winner will not be the lab with the largest compute allocation. It will be the lab with the best verification layer. The same principle determines crypto. The chain with the best fraud-proof and audit infrastructure wins institutional balances, regardless of marketing. The protocol that treats audits as a cost center becomes a liability. The protocol that treats verification as a product becomes infrastructure. OpenAI has just signaled that the same economics now apply to knowledge production. One further observation. The consultation with the Institute for Advanced Study is a double-edged signal. It indicates a desire for academic legitimacy. It also indicates internal awareness that output quality is too uneven for the academic community to accept without intervention. Legitimacy purchased through consultation is not legitimacy earned through verification. Discipline requires recording what the bulls got right. They are right about the pipeline. A system that converts thousands of open mathematical problems into a pool of candidate manuscripts, even at an 18% surface yield, has multiplied the throughput of mathematical discovery by an order of magnitude. Unverified candidates are upstream inventory. Over a decade, a faction will survive formalization and become certified theorems. The discovery rate compounds. The bottleneck is not generation; it is the human and formal check layer, which is a solvable engineering problem. They are right about test-time compute as a durable demand driver. The era of the answer model is ending. The era of the thinker model — deployed with a multi-hour inference budget — is beginning. Long-duration inference creates structural demand for H100-class clusters with sophisticated scheduling. That demand is measurable and independent of this announcement's scientific validity. My own audit experience confirms the direction: language models can summarize a transaction trace but cannot audit one, and the compute demand is real either way. They are right about the orchestration role. "AI research orchestrator" — designing hypothesis generation campaigns, selecting verification methods, managing the interface between generative and formal machinery — is the genuinely new profession. It is the intellectual equivalent of what I do when auditing a protocol: define the trust boundary, select the verification method, treat the output as a claim until the proof lands. The bulls are correct that this announcement is a job-creation event, not a replacement event. The correct response to this disclosure is not dismissal. It is calibrated skepticism with a specific demand: publish the verification coverage ratio. Until that number lands, the count of 722 is a count of liabilities, not assets. The mathematics is the message. 4,000 problems, 722 manuscripts, 372 families. One source. One tweet. One unverified milestone. The gap between the claim and the certificate is not a scandal. It is the shape of the opportunity. What happened here mirrors what happened with unaudited DeFi protocols and undisclosed exchange liabilities. Complex output arrived ahead of the verification layer. Narratives were priced faster than proofs were checked. The next decade of AI-generated knowledge — and AI-driven finance — will belong to whoever builds the machines that separate certified results from confident-but-unchecked output. The verification gap is the new alpha. The check layer is the new margin. Require the coverage ratio. Inspect the proof burden. Treat every unverified manuscript as a liability until the certificate lands. Code does not lie; intent does. The intent here is to make you count manuscripts instead of checking proofs. Count the checks instead. Silence is the only honest ledger.