Free · Open source · The SPARRING runtime
Confident isn't correct.
AI is fluent, fast, and agreeable — which is exactly how a confident wrong answer slips into a real decision. Rapier runs the SPARRING method: one AI drafts the recommendation, a different one challenges it on evidence you can check, and you get back an answer plus a plain account of what to trust in it.
SPARRING is the method.
— the engine that runs it.
The method — SPARRING
What to do, and why it works. A discipline for pressure-testing AI-assisted decisions: draft a position, then have it survive an independently-grounded challenge before anyone agrees. It's the idea — and the research behind it.
The engine — Rapier
How you actually do it. The runtime that executes the method for you: it casts the roles across different AI vendors, enforces the grounding, and returns the two-part answer — the same way, every time.
A rapier is the blade you spar with. The tool is named for the method.
◆ The mechanic
How a challenge works
01 · Draft
One AI commits to an answer
A model reads your decision and writes a real recommendation — the position it is willing to defend, not a hedge.
02 · Challenge
A different vendor attacks it
A model from another AI vendor, chosen for independence, has to find what's wrong. A challenge only counts if it points at something checkable — a source, a file, a number. Not a vibe.
03 · Resolve
You get an answer — and a trust rider
Rapier returns the recommendation, plus a separate, plain-English account of which specifics are solid and which are estimates — so you know exactly what you're acting on.
◆ Quickstart
Run it in two lines
Bring your own API keys — Rapier calls the AI vendors you configure, and nothing runs on our servers. Want the full four-phase ceremony instead of the resolver? rapier sparring.
◆ Evidence
Why trust a second AI opinion?
Because this one is built to be checked, not just believed. SPARRING was tested in a pre-registered study — and the results, including the ones that didn't go our way, are public.
Grounded, not asserted
The challenge stands on artifacts you can verify yourself — real CVEs, DOIs, RFCs — confirmed against public registries with no model in the loop.
Consequential
In testing it caught real, would-have-shipped defects on working codebases — not hypothetical ones planted for the demo.
Honest about its limits
Where extra rounds of challenge added nothing, the study says so. The method is defined by what it can prove, not what it can claim.