ARES-A-Digital-Immune-System.mp4
Project-ARES-The-AI-That-Fail.mp4
Dan started ARES with a metaphor and an instinct. The metaphor was autoimmune: a defense system that interrogates itself before declaring a threat, the way the body's T-cells and B-cells argue over what is self and what is foreign. The instinct was that cybersecurity threat analysis was being handled by tools that produced verdicts without showing their reasoning, and that an adversarial multi-agent architecture might surface what those tools concealed.
The first commitment was that ARES would not be a SaaS attempt. It would be a research project, documented openly, contributed under GPL-3.0, with the work itself as the deliverable. Dan framed the goal in his own words as "placing a stone with my name in the flow of this magnificent river." The economic model was contribution. The unit of progress was the session.
Sessions 001 through 008 established the substrate. The graph schema came first, with frozen dataclasses for nodes and edges and 110 tests covering construction, mutation guards, and serialization. The dialectical foundation was specified next: three roles (Architect, Skeptic, OracleJudge), three phases (Thesis, Antithesis, Synthesis), and a hard rule that role and phase must match or the system raises a PhaseViolationError.
The agent foundation followed. Each agent was given a typed input-output contract and a cited_fact_ids accumulator that tracked which evidence the agent had referenced. Evidence extractors were built as deterministic parsers from raw telemetry into the shared schema, with no analytical judgment of their own. The Coordinator was added to orchestrate turn-taking, validate citations against the bound evidence packet, and reject any agent message that referenced facts absent from the packet.
This was the closed-world property, and Dan committed to it as the central architectural decision of the project. Hallucinated claims would be impossible because every claim had to cite a fact_id, and every fact_id had to exist in the immutable packet. Hallucinations became schema violations.
Sessions 009 through 012 built the multi-turn debate cycles. The premise was that more turns of dialectical argument should improve verdict quality. Architect proposes a thesis, Skeptic argues against it, Architect responds, Skeptic responds, and the OracleJudge synthesizes. The new evidence rule was added to prevent agents from looping on the same facts: each turn had to introduce at least one new fact_id or the cycle terminated with NO_NEW_EVIDENCE.
The Memory Stream was built around session 007 to give agents working context across turns without violating the closed-world constraint. Redis was the planned production backend, but Phase 1 ran on in-memory implementations to keep iteration speed high.
By the end of Phase 1, the system worked end to end. Architect proposed, Skeptic objected, Oracle ruled. The pipeline ran. The question was whether the multi-turn debate produced better verdicts than a single turn would.
Phase 3 was the empirical evaluation phase, and it produced the result that reshaped the project. Dan ran structured multi-turn debate against single-turn variants across the full scenario corpus, and the multi-turn variants degraded accuracy in every configuration tested. Selective escalation, the most permissive variant, scored 72.7 percent. Single-turn scored 81.8 percent.