Software that proves itself¶
recurve turns a spec or a repo into falsifiable claims with executable probes, then burns the gap between claimed and proven down to zero — one fresh agent per cycle, behind a gate that never lies.
Install¶
git clone https://github.com/bordumb/recurve ~/tools/recurve
python3 -m pip install pyyaml
python3 ~/tools/recurve/recurve install # → ~/.local/bin/recurve
Tip
Prefer a shim, an alias, or a CI-friendly entrypoint? See the full Installation guide. The engine is Python stdlib + PyYAML — it never fights your target's toolchain.
-
Claims, not vibes
Every promise becomes a claim with an executable probe: GREEN proven, RED not yet, BROKEN can't tell. If it can't be probed, it isn't a claim — so your README, your spec, and your intent all become machine-checkable.
-
A gate that can't lie
Nothing closes on an agent's say-so. A claim goes green only when its probe passes and the fleet gate holds — and every probe keeps a trap it must fail against, so a weakened probe is caught mechanically.
-
Burn down unattended
Point it at a spec or a repo and walk away. One fresh agent per cycle takes the highest-value RED claim to GREEN, parks what it can't prove, reverts what regresses. The ledger is the only memory — bring your own agent.
Quick tour¶
# 1 · turn a spec (or an existing repo) into a gated backlog of claims
recurve init --from-prd PRD.md --suite checkout
# 2 · record what's true today — GREEN promises become a regression suite
recurve baseline checkout
# 3 · burn red → green: from your agent session (the loop skill), or headless:
recurve run # agent defaults to a bypass-permissions Claude
# 4 · audit the checks themselves — every trap must still fail, measurably:
recurve drill --fuzz # false-positive rate per probe, from generated known-bads
The primary way to run recurve is inside your agent session: invoke the
stamped loop skill and it drives gated cycles, one fresh sub-agent each.
Prefer to walk away? recurve run does the same headless. Either way the loop
believes the run-record and the gate, never the agent's word — that is the whole
point: evidence, not belief. (ways to run)
Explore¶
-
About
What recurve is, and the bet behind it.
-
Getting started
Initialize, design the claims, kick off the loop — and stop it safely.
-
Harden the checks
Who verifies the verifier? Traps re-proven, false positives measured, reference oracles compared.
-
Run data & trajectories
Every cycle is a record; the log exports as a dataset with reward provenance on every row.
-
Multi-repo & packs
Build one tree while sculpting another; distribute claims as installable packs.
-
Architecture
The model, the verification layer, the evidence trail, and the engine.