formalswarm

A confident, well-argued paragraph from an AI agent is not evidence — you cannot recompute it. FormalSwam targets three recurring failure modes: an "empty green" where a command exits zero having tested nothing, silently skipped checks with no one comparing what was asked against what was answered, and a crashed subagent whose silence gets read as consent. It turns the verdict from a judgement into a computation, making all three states mechanically visible and auditable.

The workflow separates three roles. Independent thesis writers propose solutions in isolated contexts with no shared groupthink. Adversarial critics challenge the reasoning line by line, hunting unsupported promises, omitted costs, and incompatible constraints. Seal verifiers execute the real commands you define, reading actual exit codes and case counts. A deterministic rollup then computes CONFIRM, REVISE, or INCONCLUSIVE solely from those recomputable signals — no confidence score, no model opinion. Anyone can re-run the checks and arrive at the same conclusion.

It was built for codebase change review, but the same protocol extends to document-driven problem solving: point it at a folder of design briefs, launch plans, or resource allocations, and the agents propose, the critics challenge, the verifiers recompute constraints. No repository and no Git are required. It suits engineers and teams who want a recomputable verdict in their multi-agent workflows rather than a second confident paragraph.