2.Methods

We formalize the Nanometers pipeline as a sequence of operators acting on a structured problem space. Let Π\Pidenote the set of all well-formed Erdős problem statements, and let L\mathcal{L} denote the space of natural-language mathematical arguments. The pipeline computes a composite map

Φ ⁣:Π→L×{0,1}\Phi \colon \Pi \to \mathcal{L} \times \{0, 1\}
(1)

where the first component is a candidate proof and the second is a verification bit produced by the Lean 4 type-checker. We decompose Φ\Phi into five stages.

2.1 Problem Acquisition

Definition2.1(Problem Representation).

A problem instance is a tuple π=(k,σ,τ,M)∈Π\pi = (k, \sigma, \tau, M) \in \Pi, where k∈Nk \in \mathbb{N} is the catalogue index, σ∈L\sigma \in \mathcal{L} is the problem statement, τ⊆T\tau \subseteq \mathcal{T} is a set of subject tags from a fixed taxonomy T\mathcal{T}, and M∈{open,partial,solved}M \in \{\texttt{open}, \texttt{partial}, \texttt{solved}\} is the known status.

The acquisition operator Fetch ⁣:N→Π\texttt{Fetch} \colon \mathbb{N} \to \Pi retrieves the problem instance from the Erdős Problems catalogue. Concretely, Fetch(k)\texttt{Fetch}(k) issues an HTTP request to erdosproblems.com/k and parses the response into the tuple representation via structured HTML extraction. If the request fails, the system falls back to interactive input via stdin\texttt{stdin}.

2.2 Parallel Inference

We define two inference operators, FRF_R (rigorous) and FCF_C (constructive), each backed by a distinct language model and parameterized by a fixed system prompt:

FR ⁣:Π→L,FR(π)=GPT-5.4-Pro(sR, π)F_R \colon \Pi \to \mathcal{L}, \quad F_R(\pi) = \texttt{GPT\text{-}5.4\text{-}Pro}(s_R,\, \pi)
(2)
FC ⁣:Π→L,FC(π)=Opus-4.6(sC, π)F_C \colon \Pi \to \mathcal{L}, \quad F_C(\pi) = \texttt{Opus\text{-}4.6}(s_C,\, \pi)
(3)

Here sR,sC∈Ls_R, s_C \in \mathcal{L} are fixed system prompts that bias each model toward a distinct family of proof strategies.

Definition2.2(Strategy Families).

Let S\mathcal{S} denote the space of proof strategies applicable to problems in Π\Pi. We distinguish two (not necessarily disjoint) families:
  • SR⊆S\mathcal{S}_R \subseteq \mathcal{S}: deductive strategies (induction, contradiction, compactness arguments, applications of Szemerédi regularity)
  • SC⊆S\mathcal{S}_C \subseteq \mathcal{S}: constructive strategies (probabilistic method, algebraic constructions, Lovász Local Lemma, entropy compression, container method)

The prompt sRs_R biases FRF_R toward outputs employing strategies in SR\mathcal{S}_R, and similarly sCs_C biases FCF_C toward SC\mathcal{S}_C. The families may overlap; the prompts encode a soft constraint, not a hard partition.

Both operators execute concurrently. Let tR,tC>0t_R, t_C > 0 denote the wall-clock times for FR(π)F_R(\pi) and FC(π)F_C(\pi) respectively. The parallel execution yields a total inference time of max⁡(tR,tC)\max(t_R, t_C) rather than tR+tCt_R + t_C.

Each operator includes exponential backoff with jitter: upon receiving a rate-limit response (HTTP 429), the ii-th retry is delayed by 2i+1+ξ2^{i+1} + \xi seconds, where ξ∼Uniform(0,1)\xi \sim \mathrm{Uniform}(0, 1) and i∈{0,1,2}i \in \{0, 1, 2\}, giving delays in [2,3],[4,5],[8,9][2,3], [4,5], [8,9]. If both operators fail after three attempts, the pipeline terminates with a diagnostic.

2.3 Synthesis Operator

Given the two proof attempts PR=FR(π)P_R = F_R(\pi) and PC=FC(π)P_C = F_C(\pi), the synthesis operator Σ\Sigma constructs a combined argument. We model this as a constrained optimization:

Σ(π,PR,PC)=arg⁡max⁡P∈L  μ(P∣π)s.t.supp(P)⊆supp(PR)∪supp(PC)\Sigma(\pi, P_R, P_C) = \arg\max_{P \in \mathcal{L}} \; \mu(P \mid \pi) \quad \text{s.t.} \quad \mathrm{supp}(P) \subseteq \mathrm{supp}(P_R) \cup \mathrm{supp}(P_C)
(4)

Definition2.3(Quality Measure and Support).

The quality measure μ ⁣:L×Π→[0,1]\mu \colon \mathcal{L} \times \Pi \to [0,1] assigns a score to a proof PP relative to a problem π\pi, encoding logical validity, completeness, and parsimony. The support supp(P)⊆S\mathrm{supp}(P) \subseteq \mathcal{S} is the set of proof strategies, lemmas, and techniques invoked by PP.

In practice, the optimization in (4) is approximated by a single call to Claude Opus 4.6 acting as a referee. The synthesis prompt presents both PRP_R and PCP_C along with the original problem π\pi, and requests:

  1. Identification of logical errors in PRP_R and PCP_C independently
  2. Extraction of the strongest sub-arguments from each attempt
  3. Composition into a single proof Σ(π,PR,PC)\Sigma(\pi, P_R, P_C) with all gaps closed
  4. Rendering of the result in publication-ready LaTeX

Remark2.4(Extended Thinking Budget).

The synthesis call allocates an extended thinking budget of BΣ=65,536B_\Sigma = 65{,}536 tokens, compared to BF=32,768B_F = 32{,}768 for the initial inference calls. This asymmetry reflects the observation that comparative analysis of two mathematical arguments requires deeper sustained reasoning than the generation of a single argument. When extended thinking is enabled, the temperature parameter is fixed at T=1.0T = 1.0 per API constraint.

Proposition2.5(Monotonicity of Synthesis).

Suppose μ(⋅∣π)\mu(\cdot \mid \pi) is monotone with respect to error correction: if PP is obtained from P′P' by correcting a logical error while preserving all valid sub-arguments, then μ(P∣π)≥μ(P′∣π)\mu(P \mid \pi) \geq \mu(P' \mid \pi). Then
μ(Σ(π,PR,PC)∣π)≥max⁡(μ(PR∣π), μ(PC∣π))\mu\bigl(\Sigma(\pi, P_R, P_C) \mid \pi\bigr) \geq \max\bigl(\mu(P_R \mid \pi),\, \mu(P_C \mid \pi)\bigr)

2.4 Formal Verification

The verification stage maps natural-language proofs to the type-theoretic domain. We define two operators:

  • Aristotle ⁣:L→Λ∪{⊥}\texttt{Aristotle} \colon \mathcal{L} \to \Lambda \cup \{\bot\}: the autoformalization operator, mapping natural-language proofs to Lean 4 terms (Λ\Lambda) or failure (⊥\bot)
  • Lean4 ⁣:Λ→{ok,err}\texttt{Lean4} \colon \Lambda \to \{\texttt{ok}, \texttt{err}\}: the Lean 4 type-checker

Definition2.6(Verification Predicate).

For a candidate proof P∈LP \in \mathcal{L}, define the verification predicate
V(P)={1if Aristotle(P)≠⊥   and   Lean4(Aristotle(P))=ok0otherwiseV(P) = \begin{cases} 1 & \text{if } \texttt{Aristotle}(P) \neq \bot \;\text{ and }\; \texttt{Lean4}\bigl(\texttt{Aristotle}(P)\bigr) = \texttt{ok} \\ 0 & \text{otherwise} \end{cases}
(5)

The composite pipeline then yields the output pair

Φ(π)=( Σ(π, FR(π), FC(π)),    V(Σ(π, FR(π), FC(π))) )\Phi(\pi) = \Bigl(\,\Sigma\bigl(\pi,\, F_R(\pi),\, F_C(\pi)\bigr),\;\; V\bigl(\Sigma(\pi,\, F_R(\pi),\, F_C(\pi))\bigr)\,\Bigr)
(6)

Remark2.7(Soundness).

Verification is one-sided: V(P)=1V(P) = 1 implies that PP is a valid proof of π\pi (under the assumption that the Lean 4 kernel is sound). However, V(P)=0V(P) = 0 does not imply invalidity. The autoformalization operator may return ⊥\bot for arguments that rely on informal reasoning or results not yet in Mathlib.

2.5 Artifact Generation

Each run of Φ(π)\Phi(\pi) produces a complete audit trail. Let dd be the execution date. The output directory outputs/{d}_problem{k}/ contains:

  • prompts.json: all system and user prompts sR,sCs_R, s_C with model parameters
  • gpt54pro_raw.md, opus46_raw.md: unmodified model outputs PR,PC∈LP_R, P_C \in \mathcal{L}
  • synthesis.md: the referee analysis and combined proof Σ(π,PR,PC)\Sigma(\pi, P_R, P_C)
  • proof.tex: self-contained LaTeX document with amsmath, amsthm preamble
  • formalization.lean: Lean 4 source (present when Aristotle(P)≠⊥\texttt{Aristotle}(P) \neq \bot)