Pick a scenario. The deterministic floor (git · CI · lint · SQL) grounds a blocking critic that tries to refute the artifact — and escalates to a human when nothing deterministic backs the verdict.
Adversarial loop
Blocking findings (evidence-grounded)
Artifacts
How ADRA works
Each capability runs one loop: plan → ground → generate → CRITIC → revise → escalate.
Deterministic-first — tools (git merge-base, the exact CI command, language/leak scan, test discovery, SQL probes) run first and are ground truth; the LLM cannot overturn a blocker.
Blocking critic — a two-pass red-team (deterministic hard-floor + LLM semantic) over one shared rubric; unresolved blockers escalate to a human — never silent approval.
LLM-as-judge — swap-and-average + reference anchoring to counter position/verbosity/self-preference bias.
Multi-provider — Anthropic · OpenAI · Groq · xAI · Mistral · DeepSeek · OpenRouter · Together · local Ollama, with per-role routing. Runs fully offline (deterministic mock) with no key.
Immutable provenance — every run writes an append-only record (the steps you see here).
This console drives the real ADRA engine. Provider keys live only on the server, never in the browser.