Reusable verification guidance is best treated as a scoped control aid. The current development snapshot packages skills and a shared reference library that guide specification, adequacy audits, proof construction, final audits, replay commands, and evidence reporting.
The current development snapshot packages nine skills and a shared reference library for coding agents. The live workflow selects a bundled semantics, writes the specification and its mathematical summaries, audits adequacy, constructs a proof, and requests a final proof audit. References on invariants, representation choices, and proof failures are loaded as needed. Stage outputs are claims, verification helpers, a scope document, replay commands, and audit reports. The released plugin contains eight skills, with the Rust command-line client, kprover, distributed separately to handle proof requests and evidence retrieval.
Its measured value varied by model in a frozen comparison of 31 original programs under swapped semantics. GPT-5.6 Luna accepted 19 programs with the fixed kit, compared with 15 using complete rules and 15 using generic advice. DeepSeek-V4.1-Flash accepted 16 with the kit, compared with 17 using rules and 18 using generic advice. The paired differences favored the kit for Luna and favored the comparison conditions for DeepSeek, with intervals that included zero.
We test whether the kit’s procedures improve acceptance beyond complete rules and generic advice on all 31 original programs under swapped semantics. Each program receives one attempt in each of three arms. Rules supplies the complete acceptance requirements. Generic adds advice about planning, checking, and revision. Kit instead adds a fixed historical core of procedures for scope, invariant derivation, circularity matching, residual inspection, and non-vacuity. The latter two addi- tions each contain 2,145 tokens under the same named tokenizer. Every excerpt is supplied at the start. The arms share programs, semantics, mathematical vocabulary, checks, and time limits.
5.2 FROZEN GUIDANCE WITH MATCHED CONTROLS