OVERVIEW

Verdaca docs

Verdaca is a prototype of a governed AI decision workflow: every vendor sits behind a typed port, and a multi-agent deliberation writes a JSON receipt of what it relied on, what was challenged and what each model call cost. These pages describe only what runs in the repository today. The pages under Design preview describe a planned verdaca package that is not built.

What runs today

• Deliberation with a receipt. A producer drafts from numbered evidence, a reviewer writes a counterargument before it sees the draft and then critiques it, and a synthesizer returns one of four typed outcomes: answer, clarify, abstain or escalate. Every run writes a JSON receipt with the evidence, draft, critique, diff, per-call tokens, cost and stop reason, and a SHA-256 hash of the content. The hash detects edits; the receipt is not signed.
• Two ways to run it. From a script, or as the MCP tool verdaca_deliberate in Claude Desktop. See Getting started.
• Three recorded runs (2026-10-01), all on the same question, $0.026–$0.031 each. One was called from Claude Desktop. The receipts and the server log are public in docs/receipts/. Three runs of one question are not a benchmark.
• Contract-tested seams. Memory (Mem0 ↔ Letta) and compaction (LLMLingua ↔ in-tree) swaps are proven by shared contract suites.
• Auth-first gateway. Signature, OIDC identity, replay/nonce and budget checks before any model call. The test suites verify identity against a local OIDC validator.

Not built yet

• Deliberation behind the gateway or the Teams and Slack channels. Those channels are contract-tested against recorded traffic, not validated in a live workspace.
• Signed receipts, a recorded human sign-off, and memory or adapter entries in the receipt.
• An installable verdaca package (pip install verdaca, Studio, verdaca.yaml).