SaltBench: A Referee-Gated Protocol for Measuring Method Effects in Machine-Checked Software Work
Jason Hickey
cs.SE
Sep 10, 2026 · v4
cs.LO
TL;DR
Uses a Lean 4 kernel referee, with declaration replay and an axiom audit, to score coding-agent proofs on CLEVER tasks in one of its earlier benchmark populations.
Abstract
SaltBench asks how a machine referee (a proof kernel, program verifier or withheld test suite) changes how a coding agent works. Outcomes are decided outside the agent's toolchain; a wall probed before any scored run isolates the agent from the network, reference solutions and harness; a dated freeze of predictions authorizes each run; a budget stop is a halt, never a failure. Five Rust systems components on a pinned Verus toolchain are each refereed by a withheld test suite. The arms: plain; salt-diet, also instructed to specify and verify its code (a registered reduced rendering of the method); and two arms handed the specification a priori, where the registered sign test at k=4 reached no verdict (3 of 4, p = 0.3125). salt-diet cost more on all five components, by a practical margin: no premium exceeded 2.8879x under either reading of the declared set and the three cheapest sat below 1.4x, a bound of this population and not a promise about larger ones: the premium is near 1 on the smallest components and rises with size. Versions 2 and 3 add the complete pilot matrix: 200 conditions over four models, 181 with a result of record, 16 inexpressible and 3 declared unreached at the cost cap, costed in tokens, dollars and wall time, with no verdict on the arms. Version 3 adds a correctness reading, declared post hoc and registered after every verdict it reads existed and many had been read by its author: 447 cells pass the withheld suite, 40 fail, 10 are censored at a registered budget, 4 are unscorable and 42 are read by condition. salt-diet's pass-rate interval lies wholly below plain's on 8 greenfield and 5 brownfield rows and wholly above it on 2 greenfield rows. It is descriptive, over three cells of record per condition, with no test, and says nothing about whether the method makes code more or less correct. The full record is public.
Problem
Benchmarks of coding agents can let the agent's own account or post-hoc analyst choices decide outcomes. SaltBench asks how a machine referee changes how an agent works, and what an instructed specify-and-verify method costs. The referee is a proof kernel, a program verifier or a withheld test suite.
Approach
Outcomes are decided outside the agent's toolchain under pre-registered, frozen predictions, with sandbox walls probed before scored runs. The main measurement uses five authored Rust components on a pinned Verus toolchain, comparing plain and 'salt-diet' (specify and verify) arms, with the specification either agent-written or handed over a priori. Earlier populations included S2-Lean: CLEVER tasks under Lean v4.27.0 and Mathlib. There the referee replayed declarations through the Lean kernel and restricted axioms to propext, Classical.choice and Quot.sound.
Results
Salt-diet cost more on all five components (sign test 5 of 5, p=0.0312), with premiums up to 2.89x that grow with component size. A 200-condition pilot matrix over four models reports costs but no verdict on the arms, and a post-hoc correctness pass did not separate the arms. Instrument findings from 201 scored Lean episodes and other runs are reported as transferable protocol rules.
| problem | premium, reading A | premium, reading B |
|---|
| Crc32 | 1.1610x | 1.1610x |
| FreeList | 2.8070x | 2.8879x |
| LZW | 1.3749x | 1.3749x |
| Paxos | 2.4306x | 2.2437x |
| LRU | 1.2826x | 1.1521x |
Cost premium of diet-bare over plain-bare per problem (greenfield, claude-opus-5)