ARGUS

Live demo: https://argus-lrhu.onrender.com (baked corpus + deterministic briefs; run locally with Ollama for live multi-agent deliberation)

All-source intelligence analyst workbench — a deterministic supervisor gathers from Sky (HORUS), Ocean (PHAROS), Cyber (SENTINEL), and the open-source information environment, then a multi-agent ACH deliberation panel fuses the source-rated evidence into cited intelligence briefs held to a measured evaluation bar.

Portfolio fusion point. HORUS supplies the Sky picture (ADS-B and GNSS interference), PHAROS supplies the Ocean picture (AIS maritime awareness), and SENTINEL supplies the Cyber picture (network intrusions × ATT&CK × CTI). ARGUS is the supervisor and synthesis layer: it selectively queries those read-only systems alongside its own news corpus, preserving provenance and Admiralty ratings. DELPHI is the portfolio's fifth project and the one ARGUS does not consume — it decides capacity rather than detecting events.

Status: the lean fusion supervisor, Sky/Ocean/Cyber-OSINT workers, ACH agent spine, API, overview/workbench dashboard, and narrative watch are complete (plus DSPy prompt optimization and MLX LoRA fine-tuning). What runs today and what's planned is tracked honestly in docs/ROADMAP.md; every model/agent claim lands in docs/EVAL.md with the number that survives scrutiny, not a demo cherry-pick.

What it does

Give it an analyst question — an actor, a region, an event, a timeframe — and it:

1. Collects open-source reporting on the topic (GDELT's global news index + curated world/agency RSS feeds — all free, no paid keys).
2. Structures it — extracts entities and events, deduplicates near-identical reports, and rates each source on the NATO Admiralty System (source reliability A–F × information credibility 1–6), so corroboration and provenance are first-class.
3. Routes and gathers across domains — a deterministic fusion supervisor classifies the question and wakes only the relevant mini-agents: Sky/HORUS, Ocean/PHAROS, Cyber/SENTINEL, and OSINT. Generic domain requests fuse OSINT with the matching sibling (ocean overview → OSINT + PHAROS); explicitly naming sources constrains the gather to exactly those systems (from PHAROS → PHAROS only). Broad domain synonyms keep relevant news in the fused side (ocean matches sea/maritime reporting), and thin OSINT coverage can trigger collect-on-demand. Workers make read-only HTTP calls, relevance-filter, collapse repeated same-source detector observations into one evidence episode, and return the existing rated EvidenceItem contract. Different sources remain separate as corroboration. They run no LLM; the expensive reasoning still happens once, which keeps peak memory to one local model on an M3 Pro. ARGUS_FUSION_SUPERVISOR=false restores the former flat broadcast for rollback.
4. Deliberates a brief — not a one-shot RAG summary. A multi-agent panel argues the judgment out using real intelligence tradecraft: a panel sets competing hypotheses, a lead Analyst makes the case, a Red Team attacks it (disconfirming evidence, weak sourcing, state-affiliated narratives), and an Adjudicator applies Analysis of Competing Hypotheses (favour the least-disconfirmed hypothesis) and issues a finding in the IC's estimative language (ICD 203 / Kent). The result is a structured brief — key judgments, a confidence call, the most credible alternative + what would raise it, and honest intelligence gaps — where every citation resolves to a real ingested document (fabricated citations are dropped, never shown). Built on LangGraph; runs free on a local Ollama model (Claude optional, never required), with a deterministic fallback so it always runs.
5. Watches narratives and tags influence tactics — clusters reporting into narratives, flags coordination (synchronized messaging) and influence-operations techniques (DISARM Red Framework: deepfakes, conspiracy narratives, flooding, etc.), both information-defence signals (surfaced at GET /narratives). Also detects contested-event framing divergence across sources, especially across reliability tiers, surfacing contradictions at GET /contested — all human-review signals, never automated verdicts.
6. Closes the loop on collection — a brief ends with gaps; the agent tasks the next collection by converting gaps into targeted open-source queries, ingests the results, and re-briefs on the expanded corpus (make collect-loop).
7. Shows its work before synthesis — GET /overview returns cached status, counts, and the latest item for all four lanes; POST /fusion/preview exposes the supervisor decision, per-lane counts, and fused evidence without an LLM. Every fresh brief also reports lanes_consulted and the routing reason. A named cyber actor still resolves to its contested open-reporting nation attribution (APT28 ⇄ Russia; GET /actors).
8. Speaks the standard — exports the DISARM-keyed influence-ops graph as a conformant STIX 2.1 bundle (GET /stix: techniques → attack-pattern, narratives → report, sources → identity), ingestible by OpenCTI / MISP / the ATT&CK Navigator and joinable with Sentinel's ATT&CK STIX in one object model — cyber + cognitive on the same standard.

Why it's built this way

• Honest evaluation, not a demo. The agent is measured: retrieval recall, citation accuracy (does each cited source actually exist and support the claim?), faithfulness (no ungrounded claims), and source-reliability calibration — with recorded negatives where it fails. Faithfulness and citation-support metrics are now measured via an LLM-as-judge (free/local Ollama) and cross-checked by a deterministic NLI scorer (ARGUS_NLI_ENABLED, opt-in): a cross-encoder entailment model that provides stable, independent verification alongside the LLM judge. Confidence is self-consistently calibrated (opt-in high-assurance mode via ARGUS_ASSURANCE_SAMPLES): the adjudicator is sampled K times and confidence is downgraded if the samples don't mostly agree, honest calibration over a lucky single draw. This eval harness is the point; a RAG demo that nobody graded isn't an intelligence tool. See docs/EVAL.md.
• Zero-cost, runs offline. Free data sources and free/local models only. The LLM layer is pluggable: local Ollama (auto mode; recommended: qwen2.5:14b) by default, a deterministic extractive fallback when Ollama is unreachable, and Claude (Anthropic API) as an opt-in via ARGUS_LLM_BACKEND=anthropic — never selected automatically. The product and test suite run end-to-end with no API key at all.
• Source-rated, cited, transparent. No claim without a citation; no source without a reliability rating. The Admiralty score is shown, not hidden.

Responsible use

ARGUS is strictly defensive and analytical, over public, open-source data only. It summarizes and rates open reporting to help an analyst; it does not target individuals, scrape behind authentication, profile private persons, or produce any offensive capability. Coordination signals flag patterns in public posting for human review — they are decision support, never an automated verdict. This responsible-use posture is a design constraint, not an afterthought.

Architecture

                         analyst question
                                │
                  deterministic fusion supervisor
                     route + explain (no model)
          ┌─────────────┬─────────────┬─────────────┐
          ▼             ▼             ▼             ▼
     OSINT worker   Sky / HORUS   Ocean / PHAROS  Cyber / SENTINEL
     corpus + GDELT  ADS-B + GNSS     AIS GEOINT       ATT&CK + CTI
          └─────────────┴─────────────┴─────────────┘
                                │
                    fused list[EvidenceItem]
                     ratings + provenance intact
                                │
                quick synthesis OR ACH deliberation
              hypotheses → score → analyst ⇄ red team
                         → adjudicator
                                │
                   cited, calibrated intelligence brief
                                │
        FastAPI → overview · workbench · narratives · collection

Stack

Python 3.12 · SQLAlchemy 2.0 + Alembic + Postgres · pydantic-settings · httpx + tenacity + Prefect (ingestion) · sentence-transformers + BM25 hybrid retrieval · LangGraph (multi-agent deliberation) · pluggable LLM backends (Ollama / MLX / OpenAI-compatible / Anthropic, free by default) · DSPy (prompt optimization) · MLX LoRA (self-distillation fine-tune) · FastAPI (hardened for public deploy) · React 19 + TypeScript + Vite + TanStack Query (dashboard). Mirrors Sentinel's stack and conventions so the two read as one body of work. ruff + mypy (strict) + pytest gate every change.

Quickstart

make env && conda activate argus && make install   # one-time
make up                                             # Postgres + migrations (Docker)
make ingest Q="South China Sea"                     # pull open-source reporting on a topic
make enrich                                         # entity/event extraction + Admiralty scoring
make brief Q="What happened in the South China Sea this week?"   # generate a cited brief
make eval                                           # score the agent on the gold set
make api          # read-only API + agent route on :8000

# Exercise the live read-only Sky/Ocean/Cyber workers through the supervisor:
make fusion-demo Q="Assess suspicious vessels in the Singapore Strait"

For persistent local fusion settings, copy .env.example to .env. The checked example points at the portfolio's live read-only sibling APIs; replace them with local URLs when running the four services together.

Everything works with no API key (deterministic fallback). For best results, run a local model via Ollama (ollama pull qwen2.5:14b; llama2 is too weak). To use Claude, set ARGUS_LLM_BACKEND=anthropic and ANTHROPIC_API_KEY — the auto mode never selects it.

Project context

ARGUS is one of five from-scratch, honestly-evaluated systems built as one body of work — SENTINEL (cyber), HORUS (air), PHAROS (maritime), DELPHI (infrastructure capacity) and this one (all-source fusion) — alongside smaller model-craft work such as mlx-tiny-transformer. The common thread is the evaluation discipline rather than the domain: measured baselines, pre-registered questions, and negatives written down where they fell.