Repository navigation
Rune 0.2.0 implements ADR 008 Phases 2 to 4 as far as this environment allows. cargo install rune-toolchain installs the rune binary; all seven crates are on crates.io at 0.2.0.
New: rune edit (hash-checked, Agent Booster 0.2.2 or exact replace), rune eval (independent evaluator, hidden holdout, paired statistics, explicit security gates), rune recipes, rune promote (approval bound to artifact, target, compiler, policy and contract), rune target / package / verify-artifact (cross builds, evidence classes, signed-or-unsigned evidence directory), rune cleanup (quarantine-first, never touches source), a cost ledger for rune improve, and ADR 010 (rGi as an optional supervisor, design only).
Measured (one x86_64 Linux machine): fixture kernel 3.51x (95% CI [3.511, 3.513], 40 paired samples); dropped-tail, unsafe, std::net and new-dependency candidates rejected; recipe selection p95 0.77 ms on 10,000 synthetic recipes; two live rune improve campaigns reconciled, $0.0221 total.
Not demonstrated: preregistered control study (34 kernels, 8 recipes, 111 evaluations) verdict: Insufficient — learning not demonstrated. The learned library held 3 recipes, equal to the budget, so both arms evaluated the same recipes. Also not met: RVM execution (no ABI spec), RVF packaging (published rvf-runtime cannot store an arbitrary payload), MetaHarness/AgenticOW integration, real aarch64 hardware (aarch64 and WASM runs are emulated), and large real-world workspaces. Evaluation reports are unsigned. See docs/VALIDATION.md for the full gate table.
Linux x86_64 binary built with cargo build --release --locked; checksums in SHA256SUMS.