Felipe Santibañez-Leal

Felipe Santibañez-Leal

he/him

I&F Decision Science Assoc. Manager @ Accenture | Applied Scientific Researcher

ADRA — Adversarial Dev Review Agent

Published:

A deterministic-first, adversarial-validation engine for the software lifecycle. The design premise: an LLM review that can only agree with itself is worthless, so the loop is built to disagree. Live console at adra.fasl-work.com, published on PyPI as adra, Apache-2.0.

ADRA — deterministic tools first, an adversarial critic, an LLM-as-judge, and human gates

Deterministic first, adversarial always

  • Deterministic tools run FIRST (git, the exact CI command, bundle validation, a language/leak scan, SQL probes) and their output becomes two things at once: grounding the model may not contradict, and evidence in an immutable provenance log.
  • A blocking adversarial critic tries to refute each artifact instead of blessing it; an LLM-as-judge scores with swap-and-average to cancel position bias.
  • Where nothing deterministic backs a verdict, the loop escalates to a human rather than guessing. Read-only by default, with human gates on PR create, push and merge.

Six skills (code review, PR evaluation, experiment, improve, document, decide) run the same loop over four real connectors plus an offline emulator, so the whole engine runs offline with no API key. Version 0.x on purpose: the connector-phase security controls are labelled planned, and the page says so.