Pith. sign in

REVIEW 3 major objections 4 minor 31 references

An AI Approach to Verified Production Cryptographic Libraries

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read CryptoProver claims that an AI agent, guarded by eight mechanical gates and verified with Verus, can synthesize the internal specifications and proofs that certify production cryptographic libraries without changing executable code.

desk verdict Crate-scale spec and proof synthesis is real here; the 'full verification' claim needs a caveat about a trusted-library helper that is assumed, not proved. read the letter →

arxiv 2608.00965 v1 pith:ADU6UUC7 submitted 2026-08-02 cs.CR cs.AI

classification cs.CRcs.AI
keywords formalverificationAIproofsynthesiscryptographiclibrariesVerusRusttrust-firstmechanicalgatesLLMagents
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

CryptoProver is an attempt to close the gap between AI-generated proofs and production-grade formal verification. The paper's claim is that a general-purpose coding agent, kept honest by eight mechanical gates and checked by the Verus verifier, can take a production Rust cryptographic crate plus high-level API contracts and a trusted library, and synthesize all the missing internal specifications and proof bodies needed for full functional verification—without modifying the executable code. The demonstration is concrete: curve25519-dalek, the elliptic-curve library used in Signal, was re-verified in 11.4 hours at $466.99 in API cost, and RustCrypto's previously unverified chacha20 implementation was verified against an RFC 8439 specification in a single 15-minute round. The design principle is trust-first: every route to false success is closed mechanically rather than by asking the model to behave. If true, this shifts the bottleneck of verifying deployed cryptographic code from months of expert theorem-proving labor to writing the API contracts and trusted library; the paper is careful to frame both results as existence proofs, with functional correctness against the supplied contracts as the guarantee, not constant-time or side-channel resistance.

What carries the argument

The load-bearing mechanism is the gate suite: a set of deterministic predicates computed from recorded evidence (task-start snapshots, pre/post worktree, edited files, and shell commands) that mechanically reject each known way the agent could fake success. The names encode the threats: the agent cannot leave an admit() in place (admit-count), add a new axiom_* (axiom-drift), weaken or delete a specification under proof (spec-drift), break a sibling module while fixing a target (sibling-verus), edit the harness or verifier config (tooling-drift), recover the answer from git history (git-recovery), touch frozen files (frozen-edit), or insert an assumption/external body that discharges an obli

What would settle it

Re-run the curve25519-dalek proof-and-spec synthesis on a fresh container and have an independent auditor replay the recorded per-round evidence, checking every accepted round for a construct that discharges an obligation without a proof—for example an assume(...), an #[verifier::external_body] block, or a macro-expanded admit that admit_inventory misses. If any gate-passing round contains such a construct, the trust-first claim is false; if none does, the existence result survives this check.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the remaining obstacle to AI verification of real libraries is not theorem-proving capability but process integrity. When the agent is allowed to write specifications as well as proofs, verifier acceptance alone can be faked: an earlier campaign closed 97.1% of obligations while 11 'proofs' rested on invented axioms and 5 silently broke sibling modules. CryptoProver replaces prompt-level honesty rules with eight deterministic gates—admit-count, axiom-drift, spec-drift, sibling-verus, tooling-drift, git-recovery, frozen-edit, and forbidden-construct—and accepts a round only when Verus passes, no non-axiom admits remain, and every applica

Load-bearing premise

The load-bearing premise is that the eight mechanical gates are complete enough that no unmodeled route lets a round pass Verus and all gates without actually proving the fixed contracts; the paper itself notes that the gates bound only the failure modes they encode.

Editorial extensions

If this is right

  • Given fixed API contracts and a trusted library, a production Rust crate can be functionally verified in hours and at hundreds of dollars of API cost, not months of expert labor.
  • The synthesized proofs need not resemble human proofs: the agent used 196 proof functions to the human's 235, at 48.5% as many proof lines, while still making the whole crate re-verify.
  • The verification is an independent check of shipped code: executable code never changes, and a fresh pinned-verifier container re-reports zero errors.
  • The gate discipline, not the model, is the transferable part: a second, previously unverified crate (chacha20) was verified by the same harness in one round.
  • The guarantee is functional correctness relative to the human-supplied contracts and trusted library; constant-time, side-channel, and contract-adequacy properties are explicitly out of scope.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • My inference: the architecture suggests a general recipe for any LLM proof system—enumerate the known cheats, make each one mechanically checkable, and accept only verifier-plus-gate evidence—so the eight-gate suite could be ported to other proof assistants.
  • My inference: the headline 11.4-hour figure is not end-to-end automation, since humans authored the contracts, trusted library, target decomposition, proof order, and harness; the paper explicitly excludes that setup effort from the comparison.
  • My inference: a third-crate test on an adversarially selected crate without a human reference, run under the same fresh-container protocol, would be the natural check of whether the existence results generalize to average performance.
  • My inference: the residual risk that most limits the trust story is the gate suite's own completeness—the paper concedes it 'does not rule out an unmodeled bypass'—so a red-team search for a construct that passes all eight gates without proving the contract is the sharpest open question.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. The paper presents CryptoProver, a driver loop around an LLM coding agent that writes internal specifications and proof bodies for Rust crates, with acceptance enforced by Verus and a set of eight mechanical 'gates.' The two headline results are (i) synthesis of all internal specifications and proofs for curve25519-dalek in 11.4 elapsed hours at $466.99 in recorded API cost, with a whole-crate Verus run reporting 2,031 checks and zero errors (replicated with a second model at 2,114 checks), and (ii) verification of RustCrypto's chacha20 v0.10.1 portable backend against a human-authored RFC 8439 specification. The paper claims this is achieved without changing executable code, and that the design prevents specification weakening, invented axioms, sibling-module breakage, and recovery of the reference proof from git history. The evidence is supported by a public artifact, machine-checked acceptance, and unusually candid disclosure of run-specific limitations and threat-model boundaries.

Significance. If the central claims hold, this is a substantial advance in LLM-based formal verification: it moves from function-level or module-level proof synthesis to crate-scale proof-and-spec synthesis with machine-checked acceptance, and it explicitly quantifies cost and trusted inputs. The strengths of the paper are the whole-crate Verus evidence, the public artifact, the second-model replication, and the honest listing of threats in Section 5.3 and the appendices. The main weaknesses are that the strongest headline statements outrun the disclosed setup: the input boundary omits several human-supplied components that are necessary for the 11.4-hour figure, and the trusted library contains at least one unproved `assume(false)` helper on which the Montgomery proofs depend. Neither issue invalidates the core system, but both must be fixed in the claims before the paper can be accepted as stated.

major comments (3)
  1. [Abstract and §1 vs. §5.3] The abstract and §1 say CryptoProver synthesizes the proof 'given just the top-level API contracts and the trusted library.' §5.3, however, states that the 11.4-hour figure measures agent elapsed time after the contracts, the trusted library, the specification vocabulary, the target decomposition, the proof order, and the harness had been supplied. These additional human inputs are load-bearing for the automation claim: the agent did not derive the decomposition or the proof order, and internal specification vocabulary was fixed. The wording should be corrected in the abstract and introduction, for example by saying 'given API contracts, a fixed trusted library, a fixed specification vocabulary, and a target decomposition.'
  2. [§4.1 and §G.2] The claim of 'no unresolved proof obligations' is not logically established as stated. Appendix G.2 reports that the generated Montgomery proofs call the pre-existing trusted-library helper `lemma_u128_shl_is_mul`, whose body is `assume(false)`, documented as pending vstd support. A body of `assume(false)` discharges the ensures without proof. This is an unresolved obligation hidden inside the frozen trusted base, and it is invisible to `forbidden-construct`, which counts `assume` only in editable files (Appendix A). Unless the helper's conclusion is independently proved, the 2,031-check acceptance establishes the API contracts only relative to an unproved arithmetic fact. The paper should either prove this helper or explicitly qualify the headline 'full functional verification' claim in the abstract and main text as holding modulo this trusted-library placeholder.
  3. [§4.1 and §G.1] The term 'independent proof' is stronger than what the experiment supports. The synthesis starts from a stripped tree of the human reference proof: proof bodies and internal lemma statements are removed, but the API contracts, the internal specification vocabulary, and the file/module structure remain fixed. Moreover, §G.1 states that the editable files retained roughly 5,800 lines of comments inherited from the verified source, including three proof-strategy comments. The agent therefore had substantial guidance from the human proof architecture. This does not negate the contribution, but the 'independent' claim in §1 should be qualified to mean 'not retrieved from git history, and with newly synthesized proof bodies and intermediate statements,' rather than implying a proof effort independent of the human reference.
minor comments (4)
  1. [Abstract and §4.1] The abstract says CryptoProver 'verifies RustCrypto's previously unverified chacha20 implementation,' but §4.1 states that the verification covers the portable backend, not the SIMD backends. The abstract should include this scope limitation, since a reader could reasonably assume the entire library is verified.
  2. [Appendix A, forbidden-construct] The `forbidden-construct` gate counts `assume(...)` and `#[verifier::external_body]` only across editable files. The main text should explicitly state that the trusted library is allowed to contain such constructs, and that this is a deliberate part of the trust base. Currently the distinction is easy to miss, and it is directly relevant to the `assume(false)` helper in §G.2.
  3. [Figure 3] The annotations in the left panel (e.g., '166 154 113 69 0 ...') are hard to interpret. A brief legend or a note explaining whether these are per-sweep or per-attempt error counts would improve readability.
  4. [§3.6, git-recovery] The `git-recovery` gate is described as a blocklist of command forms. Because command-name matching can be bypassed by unusual invocations, the real guarantee comes from the sandbox excluding the reference proof. The text already notes this, but it would be clearer to state upfront that the sandbox, not the command blocklist, is the primary protection.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the acceptance criterion is an external verifier (Verus) checking fixed human-supplied API contracts, and no load-bearing claim reduces to its own inputs by construction.

full rationale

The paper's derivation chain is: fixed API contracts + fixed trusted library + agent-synthesized internal specifications/proofs, with whole-crate Verus acceptance plus mechanical gates. The theorem proved is the human-supplied contract, checked by an external SMT-backed verifier, not a quantity derived from the agent's own output. No fitted parameter, normalization, or specification choice is back-solved from the target result. The gates are process-integrity checks and are explicitly scoped: 'The gate suite bounds only the failure modes it encodes; it does not rule out an unmodeled bypass' (Section 5). This is a stated limitation, not a circularity. The human reference proof is used as a benchmark, not as an input: isolation blocks reference-proof retrieval, and the agent's proofs must pass fresh-container Verus checks. Self-citations (VeriStruct, Clover, Lemur, CryptOpt) appear only as background on fallible specifications and related systems; none is load-bearing for the central claim. Appendix G.2 reports that the generated Montgomery proofs call the pre-existing trusted-library helper lemma_u128_shl_is_mul with an assume(false) body 'documented as pending vstd support.' This is a real soundness caveat about the trusted base, but it is not a circular derivation: the paper does not claim that helper was proved by the agent, and it explicitly frames the guarantee as relative to the trusted library. The verification result is therefore weaker than an absolute 'full functional correctness' claim, but it is not equivalent to its inputs by construction. Consequently, no circular step is present.

Assumptions & free parameters 0 free parameters · 5 assumptions · 0 invented entities

The central claim rests on human-supplied contracts, a human-supplied trusted library, toolchain soundness, and gate coverage; no numeric constants are fitted to data. Solver rlimit attributes are resource budgets, not free parameters in the scientific sense.

assumptions (5)
  • domain assumption The trusted library (field specs, field/common arithmetic facts, 48 axiom_* lemmas, vstd) is correct
    Section 2.2 defines the trusted library as assumed correct or proved elsewhere; the verification inherits these assumptions.
  • domain assumption The human-authored API contracts and RFC 8439 formal spec correctly capture intended functional behavior
    Section 4: contracts and spec are fixed human inputs; Section 5.3: the result inherits contract adequacy as an unverified premise.
  • standard math Verus and Z3 are sound for the checked obligations
    The agent's proofs are accepted only via whole-crate Verus runs (Section 3.4); soundness of the toolchain is assumed.
  • domain assumption The eight mechanical gates have no unmodeled bypass
    Section 5: 'it does not rule out an unmodeled bypass'; the trust-first argument depends on gate coverage of all false-success routes.
  • domain assumption The shared helper lemma_u128_shl_is_mul (body assume(false), pending vstd support) is sound
    Section G.2 says generated Montgomery proofs call this trusted-library helper, whose body is a documented pending assumption; the reference proof calls the same helper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An AI Approach to Verified Production Cryptographic Libraries." pith.science (2026). https://pith.science/paper/ADU6UUC7

@misc{pith2026260800965,
  author       = {Pith},
  title        = {Pith review of: An AI Approach to Verified Production Cryptographic Libraries},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ADU6UUC7}},
  note         = {Machine review of arXiv:2608.00965}
}
read the original abstract

Cryptographic code is critical infrastructure that must be correct, yet formally verifying production libraries remains difficult. Existing language-model proof systems solve isolated obligations with specifications and premises already given, leaving production-library verification unresolved. We present CryptoProver, an AI-based system that synthesizes internal specifications and Verus-checked proofs from high-level API contracts. Without changing executable code, CryptoProver constructs a new independent proof of curve25519-dalek and verifies RustCrypto's previously unverified chacha20 implementation against an RFC 8439 specification. These cryptographic lineages underpin deployed systems including Signal and Shadowsocks; Signal has an estimated 218M global downloads. The independent, human-led curve25519-dalek verification was developed publicly over eight months by five main contributors. Given the API contracts and a fixed trusted library of field specifications, arithmetic facts, axioms, and vstd, CryptoProver synthesizes the internal specifications and proofs in 11.4 hours with USD 466.99 in recorded API cost. CryptoProver follows a trust-first design principle: mechanical gates reject specification weakening, invented axioms, and cross-module breakage, while isolation blocks reference proof retrieval, including from git history.

Figures

Figures reproduced from arXiv: 2608.00965 by the authors.

Figure 1
Figure 1. The outer loop sweeps the configured targets in order; the inner loop iterates rounds within one attempt. verified counterexample ends the attempt as FALSE_CONTRACT. A later attempt may retry the same target, seeded with per-target failure memory that records declaration-level errors from earlier attempts; a prior NEEDS_DECOMP earns the retry a larger round and wall-clock budget. When an attempt ends, the driver dis… view at source ↗
Figure 2
Figure 2. The artifact layers of a verified crate. Shaded layers are human-supplied fixed input; dash-outlined layers are agent-synthesized. Proof-and-spec synthesis targets the intermediate internal specifications and proof bodies above the fixed trusted library. zero errors at the default resource limit; the final tree contained no unresolved proof obligations, no executable-code changes, and exactly 48 axioms, all already … view at source ↗
Figure 3
Figure 3. Error trajectories on the proof-and-spec synthesis task. The proof-and-spec synthesis [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Agent and human proof functions by source file. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: No-hints run dynamics: (a) minutes of proving per closed admit; (b) cumulative admits [PITH_FULL_IMAGE:figures/full_fig_p019_5.png]
Figure 6
Figure 6. Figure 6: Start and audited end states of the proof-and-spec synthesis run. [PITH_FULL_IMAGE:figures/full_fig_p024_6.png]
Figure 7
Figure 7. Figure 7: Share of elapsed time spent in the verifier for the proof-and-spec synthesis comparison. [PITH_FULL_IMAGE:figures/full_fig_p025_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 14 canonical work pages

  1. [1]

    Debian Security Advisory

    Debian Security Team.DSA-1571-1: New OpenSSL Packages Fix Predictable Random Number Generator. Debian Security Advisory. Accessed 2026-07-13. 2008.url: https://lists. debian.org/debian-security-announce/2008/msg00152.html

  2. [2]

    Lorch, Shuai Lu, Fan Yang, Ziqiao Zhou, and Shan Lu.AutoVerus: Automated Proof Generation for Rust Code

    Chenyuan Yang, Xuheng Li, Md Rakib Hossain Misu, Jianan Yao, Weidong Cui, Yeyun Gong, Chris Hawblitzel, Shuvendu Lahiri, Jacob R. Lorch, Shuai Lu, Fan Yang, Ziqiao Zhou, and Shan Lu.AutoVerus: Automated Proof Generation for Rust Code. OOPSLA 2025. 2024.doi: 10.1145/3763174. arXiv:2409.13082

  3. [3]

    LMPL 2025

    Si Cheng Zhong and Xujie Si.Towards Repository-Level Program Verification with Large Language Models. LMPL 2025. 2025. arXiv:2509.25197

  4. [4]

    Yuwei Liu, Xinyi Wan, Yanhao Wang, Minghua Wang, Lin Huang, and Tao Wei.KVerus: Scalable and Resilient Formal Verification Proof Generation for Rust Code. 2026. arXiv: 2605.03822

  5. [5]

    2025.doi:10.48550/arXiv.2510.25015

    Chuyue Sun, Yican Sun, Daneshvar Amrollahi, Ethan Zhang, Shuvendu Lahiri, Shan Lu, David Dill, and Clark Barrett.VeriStruct: AI-assisted Automated Verification of Data-Structure Modules in Verus. 2025.doi:10.48550/arXiv.2510.25015. arXiv:2510.25015

  6. [6]

    Verus: Verifying Rust Programs using Linear Ghost Types

    Andrea Lattuada, Travis Hance, Chanhee Cho, Matthias Brun, Isitha Subasinghe, Yi Zhou, Jon Howell, Bryan Parno, and Chris Hawblitzel. “Verus: Verifying Rust Programs using Linear Ghost Types”. In:Proc. ACM Program. Lang. (OOPSLA). 2023. arXiv:2303.05491

  7. [7]

    GitHub repository

    Beneficial AI Foundation.curve25519-dalek: independent Verus verification fork. GitHub repository. 2026.url: https://github.com/Beneficial-AI-Foundation/dalek-lite/ pull/774

  8. [8]

    GitHub repository

    Signal Messenger, LLC.libsignal v0.96.3. GitHub repository. 2026.url:https://github. com/signalapp/libsignal/tree/v0.96.3

Show all 31 references
  1. [9]

    Sensor Tower

    Kara Lee.All Signals Point Up for App’s Downloads and MAUs. Sensor Tower. Accessed 2026-07-13. 2025.url: https://sensortower.com/blog/all-signals-point-up-for- apps-downloads-and-maus

  2. [10]

    GitHub repository

    Shadowsocks Contributors.shadowsocks-rust. GitHub repository. 2026.url:https://github. com/shadowsocks/shadowsocks-rust/tree/c88b519

  3. [11]

    Z3: An Efficient SMT Solver

    Leonardo de Moura and Nikolaj Bjørner. “Z3: An Efficient SMT Solver”. In:Tools and Algorithms for the Construction and Analysis of Systems (TACAS). 2008, pp. 337–340

  4. [12]

    Lorch, and Shan Lu.VeruSAGE: A Study of Agent-Based Verification for Rust Systems

    Chenyuan Yang, Natalie Neamtu, Chris Hawblitzel, Jacob R. Lorch, and Shan Lu.VeruSAGE: A Study of Agent-Based Verification for Rust Systems. 2025. arXiv:2512.18436

  5. [13]

    Lost in the Middle: How Language Models Use Long Contexts

    Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. “Lost in the Middle: How Language Models Use Long Contexts”. In: Transactions of the Association for Computational Linguistics(2024). TACL 2024. arXiv: 2307.03172

  6. [14]

    HACL*: A Verified Modern Cryptographic Library

    Jean Karim Zinzindohoué, Karthikeyan Bhargavan, Jonathan Protzenko, and Benjamin Beurdouche. “HACL*: A Verified Modern Cryptographic Library”. In:Proc. ACM CCS. 2017, pp. 1789–1806.doi:10.1145/3133956.3134043. 12

  7. [15]

    EverCrypt: A Fast, Verified, Cross-Platform Cryptographic Provider

    Jonathan Protzenko, Bryan Parno, Aymeric Fromherz, Chris Hawblitzel, Marina Polubelova, Karthikeyan Bhargavan, Benjamin Beurdouche, Joonwon Choi, Antoine Delignat-Lavaud, Cédric Fournet, Natalia Kulatova, Tahina Ramananandro, Aseem Rastogi, Nikhil Swamy, Christoph M. Winterste...

  8. [16]

    GitHub repository

    chacha20-verus: Verus-Verified Fork of RustCrypto’s chacha20. GitHub repository. 2026.url: https://github.com/oliversssf2/chacha20-verus

  9. [17]

    Enhancing LLM-Based Proof Synthesis for Rust Programs via Semantic Chunking and Hierarchical Context Expansion

    Yuchen Zhang, Cheng Wen, Zhiwu Xu, Dugang Liu, Jialun Cao, Yuwei Liu, Shengchao Qin, and Cong Tian. “Enhancing LLM-Based Proof Synthesis for Rust Programs via Semantic Chunking and Hierarchical Context Expansion”. In:Theoretical Aspects of Software Engineering. Springer Nature...

  10. [18]

    Clover: Closed-Loop Verifiable Code Generation

    Chuyue Sun, Ying Sheng, Oded Padon, and Clark Barrett. “Clover: Closed-Loop Verifiable Code Generation”. In:Proc. iFM. 2024. arXiv:2310.17807

  11. [19]

    Zhe Ye, Zhengxu Yan, Jingxuan He, Timothe Kasriel, Kaiyu Yang, and Dawn Song.VERINA: Benchmarking Verifiable Code Generation. 2025. arXiv:2505.23135

  12. [20]

    Lemur: Integrating Large Language Models in Automated Program Verification

    Haoze Wu, Clark Barrett, and Nina Narodytska. “Lemur: Integrating Large Language Models in Automated Program Verification”. In:Proc. ICLR. 2024. arXiv:2310.04870

  13. [21]

    Shubham Agarwal, Alexander Krentsel, Shu Liu, Mert Cemri, Audrey Cheng, Rui Meng, Tomas Pfister, Chun-Liang Li, Sylvia Ratnasamy, Aditya Parameswaran, Matei Zaharia, Ion Stoica, and Mohsen Lesani.Inductive Deductive Synthesis: Enabling AI to Generate Formally Verified Systems....

  14. [22]

    Verified Correctness and Security of mbedTLS HMAC-DRBG

    Katherine Q. Ye, Matthew Green, Naphat Sanguansin, Lennart Beringer, Adam Petcher, and Andrew W. Appel. “Verified Correctness and Security of mbedTLS HMAC-DRBG”. In:Proc. ACM CCS. 2017.doi:10.1145/3133956.3133974. arXiv:1708.08542

  15. [23]

    2019.doi:10.48550/arXiv.1904.04606

    José Bacelar Almeida, Manuel Barbosa, Gilles Barthe, Benjamin Grégoire, Adrien Koutsos, Vincent Laporte, Tiago Oliveira, and Pierre-Yves Strub.The Last Mile: High-Assurance and High-Speed Cryptographic Implementations. 2019.doi:10.48550/arXiv.1904.04606. arXiv: 1904.04606

  16. [24]

    2023.doi:10.48550/arXiv.2211.10665

    Joel Kuepper, Andres Erbsen, Jason Gross, Owen Conoly, Chuyue Sun, Samuel Tian, David Wu, Adam Chlipala, Chitchanok Chuengsatiansup, Daniel Genkin, Markus Wagner, and Yuval Yarom.CryptOpt: Verified Compilation with Randomized Program Search for Cryptographic Primitives. 2023.d...

  17. [25]

    Natalia Klaus, Juan Conejero, and Palina Tolmach.A Rust-to-Lean Verification Pipeline with AI Provers: An Experience Report. 2026. arXiv:2605.30106

  18. [26]

    Max Tan.Automating Formal Verification with Reinforcement Learning and Recursive Infer- ence. 2026. arXiv:2605.30914

  19. [27]

    Z. Z. Ren, Zhihong Shao, Junxiao Song, Huajian Xin, Haocheng Wang, Wanjia Zhao, Liyue Zhang, Zhe Fu, Qihao Zhu, et al.DeepSeek-Prover-V2: Advancing Formal Mathematical Reasoning via Reinforcement Learning for Subgoal Decomposition. 2025. arXiv:2504.21801. 13

  20. [28]

    LeanDojo: Theorem Proving with Retrieval- Augmented Language Models

    Kaiyu Yang, Aidan M. Swope, Alex Gu, Rahul Chalamala, Peiyang Song, Shixing Yu, Saad Godil, Ryan Prenger, and Anima Anandkumar. “LeanDojo: Theorem Proving with Retrieval- Augmented Language Models”. In:Proc. NeurIPS Datasets and Benchmarks. 2023. arXiv: 2306.15626

  21. [29]

    Pranjal Aggarwal, Bryan Parno, Sean Welleck, et al.AlphaVerus: Bootstrapping Formally Verified Code Generation through Self-Improving Translation and Treefinement. 2024. arXiv: 2412.06176

  22. [30]

    Sergiu Bursuc, Theodore Ehrenborg, Shaowei Lin, Lacramioara Astefanoaei, Ionel Emilian Chiosa, Jure Kukovec, Alok Singh, Oliver Butterley, Adem Bizid, Quinn Dougherty, Miranda Zhao, Max Tan, and Max Tegmark.A benchmark for vericoding: formally verified program synthesis. 2025....

  23. [31]

    Bounds on Multiprocessing Timing Anomalies

    R. L. Graham. “Bounds on Multiprocessing Timing Anomalies”. In:SIAM Journal on Applied Mathematics17.2 (1969), pp. 416–429.doi:10.1137/0117039. A Gate Definitions This appendix defines each gate formally. LetS0 be the harness-owned snapshot recorded at task start. Each round t...

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.