Skip to content

0093: FlashProve as a pinned external proof backend

Date: 2026-11-03. Status: owner-requested; local verifier and H200 proving accepted. Merged in PR #182. Numbered 0093 because the Slurm profile decision merged first as 0092.

The owner asked for support of Anthropic's FlashProve prover (announcement, repository). FlashProve proves that a committed Mixture-of-Experts transformer, run under greedy decoding on a request of up to 1,024 tokens, chose each checked token. Its zkSNARK is hash-based. It hides the weights except for a declared list, and its protocol soundness is proved in Lean. The C++ verifier and the CUDA prover are tested against that protocol; neither is itself proved.

Scope. Hallmark adds FlashProve the way it added SASH: pinned upstream source plus command-line adapters, and no copy in the tree. The pin is commit 9e2bdd8ff5bf9c62f18263d7dac189b0400b260c (v0.16.0, Apache-2.0), as a 158 MB codeload archive with a fixed SHA-256 and byte count, recorded in source.json. No upstream source file is changed. This adds a backend; it changes no existing contract, model, threshold, enrollment or identity. FlashProve cannot prove Hallmark's dense TinyLlama case because it supports MoE models only. The LLM path stays SASH.

Verifier. make verify-flashprove downloads the archive, builds cp_verify on the host and checks the two stored upstream proofs through upstream's integrations/check_proof.py. Every value comes from the caller: commitment ID, checked range, next token, configuration and request. A check passes only under upstream's own three-condition rule (exit 0, result: outcome=verified second to last, matching fields), restricted further to claim=exact-match and the pinned verifier source digest 93afdaf592393a54ebd17b681d1c03295eb883f0161ff0e6ab872a3b39939595. On Linux the build is upstream verifier/build.sh, unchanged. On macOS that script cannot run, because it needs GNU stat, mv -T and ELF linker flags. Hallmark then compiles the same unchanged cp_verify.cpp with Apple clang, using the same macros and the same source digest computed the way upstream computes it. It substitutes libc++ hardening for _GLIBCXX_ASSERTIONS and drops the ELF-only link flags. The compiler and flags differ from upstream's tested build, so the Mac binary is labelled apple-clang in every result.

Acceptance: both stored proofs are accepted, and all 99 altered sets from upstream's three rejection tables are rejected with the named check. A pass shows only that the verifier and these synthetic-weight files fit together.

Prover. cp_commit/cp_prove need an sm_90 Hopper GPU, CUDA 13.2 Update 2 and an x86-64 CPU with SHA extensions. HF Jobs offers no H100; its H200 is also sm_90 (Hopper), and the recorded HF H200 driver (580.178) matches the 580 that upstream ran under. The owner approved one run on an H200 with a USD 13 ceiling. The agent-selected controller cap is USD 3.50. At live prices, with HF's 600 s startup window counted per phase, the worst case is USD 3.27: cpu-xl build 3,600 s, H200 900 s, one attempt each, no retries or top-ups. The owner reported a USD 31 balance. Only upstream's synthetic small kit (target_small_f2_roomy) is proved. Llama 4 Maverick weights are under Meta's gated licence, which the run authority excludes (gated_terms: false).

make e2e-flashprove first runs the free local verifier suite, then two jobs. The build job runs upstream verifier/build.sh and prover/build.sh with upstream's documented pip install of CUDA 13.2.2, as an ordinary account. Its link stub comes from the pinned nvidia/cuda 13.2.1 image, because the build host has no driver. The prove job runs upstream run_target_example.sh with the retained build. It returns only public files: config/, ivk/, the request, the token and the proof. The secret key directory isk/ never leaves the worker. The CUDA wheels come from PyPI without hash pins, as upstream's instructions do; their installed versions are retained.

Caller acceptance for the fresh proof: - the worker's configuration and request are byte-equal to the pinned kit files; - the commitment ID is the one cp_commit printed, which is valid for a commitment the caller made itself; - the Mac verifier accepts under the same rule as the stored proofs; - six altered copies are rejected: two proof bits, the next token, a request token, a key bit and the commitment ID.

Not claimed. No served model, Maverick proof, weight hiding, backend qualification, security level, or soundness of the C++ verifier or CUDA prover. The bits= figures cp_verify prints are its own accounting, not a security level. The Lean verifier (tools/lean-verify) is not run.

Results

Numeric evidence.

Local verifier (Mac, Apple clang build): make verify-flashprove took 97.6 s, including the archive download and a 22 s compile. Both stored proofs were accepted through check_proof.py (small 1.45 s; full size, 1,001 tokens, 7.99 s), and all 99 altered sets were rejected. A later re-timing of raw cp_verify (five calls each, 2026-11-04) gave medians of 0.62 s and 7.88 s, with a peak of 4.23 GB for the full-size proof (comparison).

Build (cpu-xl job 6aea16ea0e0338d1bb7f5fb6): 1,520.5 s, every unit compiled by nvcc V13.2.86, with a run path to the pip toolkit.

First prove attempt failed and is retained (runs/flashprove-gpu-g9nv07j3, job 6aea1d1f0e0338d1bb7f630a, 80 s). The worker account's env -i dropped the image PATH, so nvidia-smi (under /usr/local/nvidia/bin) was not found, and the job exited 127 before any commit or proof. Fix: keep the image PATH and LD_LIBRARY_PATH for that account, and reuse the retained build after rechecking its hashes (--build-run). One corrected prove job followed.

Prove (H200 job 6aea1e650e0338d1bb7f6381, 74.4 s billed running, driver 580.178.04, Xeon Platinum 8581C with SHA-NI): upstream's commit, prove and verify example took 22.3 s wall time at 0.9 GB peak RSS. Commitment 64fbdc19…e81e17; 48 checked tokens plus next token 284; proof 4,145,980 bytes. Linux cp_verify printed YES. The Mac verifier accepted the proof under the caller rule in 0.97 s. All six altered copies were rejected (proof-refused ×3, reply-differs, key-refused, model-statement). No isk/ file was retained.

Spend: the computed upper bounds total USD 0.78 (build 0.44, failed prove 0.17, prove 0.17), within the USD 3.50 cap and the owner's USD 13. These are not an invoice. All buckets were deleted after hash-checked retention.