AGON is a Rust conflict-intelligence engine from TACITUS. It turns messy human text into a typed, evidence-backed conflict scene: actors, claims, commitments, contradictions, friction patterns, source spans, review questions, and provenance.
- What it is: a perception stack for conflict, built like infrastructure rather than a chat prompt.
- What it proves today: a 17-crate Rust workspace, Cloud Run demo, Vertex AI extraction path, 70 green tests, evidence-span verification, contradiction detection, friction matrices, and named pattern detection.
- What it is not yet: a finished legal/mediation product or a substitute for expert judgment.
- License status: public source preview; final open-source license is pending, as reflected in LICENSE.
- Best place to start: run the live endpoint examples below, then read ARCHITECTURE.md and ROADMAP.md.
█████╗ ██████╗ ██████╗ ███╗ ██╗
██╔══██╗██╔════╝ ██╔═══██╗████╗ ██║
███████║██║ ███╗██║ ██║██╔██╗ ██║
██╔══██║██║ ██║██║ ██║██║╚██╗██║
██║ ██║╚██████╔╝╚██████╔╝██║ ╚████║
╚═╝ ╚═╝ ╚═════╝ ╚═════╝ ╚═╝ ╚═══╝
conflict is legible.
perception is sovereign.
A perception engine for human conflict. AGON reads messy human text — emails, transcripts, depositions, board minutes, chat logs — and returns a typed, evidence-backed picture of the conflict inside it: who said what, what was promised, what changed, where the contradictions are, what patterns are present. Same shape as a self-driving stack — sensors, encoders, extraction, tracking, scene, calibration, provenance — applied to language.
Status: v0.1.4 live · 17-crate Rust workspace · Cloud Run + Vertex AI · 70 tests green · 3 named patterns (DARVO + Anchoring + Conspicuous Absence) · public source preview, license pending
Built by TACITUS.
Generic LLMs are good at summarizing text. They are bad at:
- Naming the move. They tell you "there's tension"; they don't tell you "this is a textbook DARVO".
- Anchoring claims to source. They paraphrase. AGON requires every primitive to cite an exact verbatim span — and verifies it.
- Tracking commitments through time. They forget what was promised three turns ago. AGON keeps a state machine:
made → confirmed → contested → broken. - Knowing when to abstain. They are confidently wrong. AGON has calibrated confidence and conformal-prediction abstention (on the roadmap).
- Being auditable. They are a black box. AGON's output is a typed DAG with content-hash provenance for every node.
AGON is not a chatbot. It is infrastructure for conflict vision — built like a perception stack, not a prompt template.
RAW TEXT
│
▼
┌───────────────────────────────────────────────────────────────┐
│ L1 SENSORS deterministic Rust │
│ canonical text · segmentation · quoted-speech FSM · │
│ speaker turns · time expressions · lexical features │
│ (aco-text · aco-time · aco-lex) │
├───────────────────────────────────────────────────────────────┤
│ L2 ENCODERS ort 2.x (ONNX Runtime) │
│ BGE-M3 embeddings · DeBERTa-v3-large NLI · fastcoref │
│ (aco-encode) — PROMPT 05 │
├───────────────────────────────────────────────────────────────┤
│ L3 EXTRACTION Vertex Gemini 2.5 Flash + Pro │
│ schema-constrained ACO primitives: │
│ Actor · Claim · Interest · Constraint · Leverage · │
│ Commitment · Event · Narrative · Contradiction │
│ (aco-extract · aco-llm) │
├───────────────────────────────────────────────────────────────┤
│ L4 TRACKING deterministic Rust │
│ cross-doc actor resolution · commitment state machine · │
│ Allen-13 temporal logic · evidence-span verification │
│ (aco-fuse · aco-temporal) │
├───────────────────────────────────────────────────────────────┤
│ L5 SCENE hybrid │
│ friction matrix · pattern library │
│ DARVO · anchoring · scope creep · conspicuous absence · │
│ coalition · power dynamics │
│ (aco-patterns) — PROMPT 09 │
├───────────────────────────────────────────────────────────────┤
│ L6 CALIBRATION deterministic Rust │
│ per-detector temperature/isotonic · stacked LR · │
│ conformal prediction for abstention │
│ (aco-score) — PROMPT 10 │
├───────────────────────────────────────────────────────────────┤
│ L7 PROVENANCE deterministic Rust │
│ typed lineage DAG · Merkle audit log · signed records · │
│ JSON-LD + Markdown export │
│ (aco-prov) — PROMPT 11 │
├───────────────────────────────────────────────────────────────┤
│ L8 DECISION Axum + SSE │
│ quality gates · review questions · streaming workbench │
│ (aco-server) │
└───────────────────────────────────────────────────────────────┘
Each layer has a typed contract. Each layer is independently testable. No single model is asked to do everything. The chassis is Rust. ML models are interchangeable passengers behind typed traits.
Paste a multi-turn dispute. Get back a structured perception.
Sam (Mon 09:14): So we're agreed — you own the Q4 launch deck content,
I handle design. Lock it in by Thursday?
Alex (Mon 09:47): Sounds good. I'll pick it up after the Jenkins pitch.
Alex (Thu 09:02): I never said I'd own it. Just help.
Sam (Thu 09:15): That's not what we discussed. We don't have time to
relitigate this — the launch is Monday.
Alex (Thu 09:18): You're putting words in my mouth. You said you'd own
the content if I helped with design.
- 2 actors —
actor_sam,actor_alex - 1 contested commitment — "own the Q4 launch deck content" · state=
contested· confidence 0.76 - 1 escalation loop around
actor_alex· confidence 0.71 - 10 contradictions with evidence spans
- 5 speaker turns detected pre-extraction
- 33/33 evidence quotes verified against canonical source
- friction matrix: Sam ↔ Alex heat 100/100, reasons include
commitment_contested,pattern: defensiveness,pattern: criticism,pattern: stonewalling,escalation_signal - review questions surfaced: "What exact words created or limited the alleged commitment?", "Which contradiction is material to the decision?"
The friction matrix and force-directed actor/claim graph render in a dark-mode workbench at https://agon-dev-tbryoen6qa-uc.a.run.app (user AGON / pass AGON).
bash scripts/agon-up.sh # start (~30 s, then ~$3–8/day active)
bash scripts/agon-down.sh # stop (~10 s, then ~$0.20–0.50/day idle)
bash scripts/agon-status.sh # status
bash scripts/agon-nuke.sh # terraform destroy (DATA LOSS — use with care)PowerShell wrappers: scripts/*.ps1. Full operator's guide: docs/AGON_GUIDE.md.
A typical few-days-of-testing cycle costs under $30 total. Your GCP startup credit covers it many times over.
These are locked. If something downstream conflicts, it loses.
- Rust chassis, ML passengers. No Python sidecar. Every model behind a Rust trait.
- JSON Schema is the source of truth.
tacitus-contractsis the only place primitives are defined. Rust types live alongside; Python and TypeScript regenerate from the same schemas. - Evidence-span quad form. Every claim-bearing primitive carries
(segment_id, canonical_offsets, raw_offsets, verbatim_quote, quote_hash, normalization_version). Non-negotiable. This is what makes a primitive auditable instead of plausible. - Calibration is mandatory. Every detector emits raw signal; the calibration registry converts to probability. LLM verbalized confidence is a feature, never a probability.
- Per-doc perception, then cross-doc fusion. Long-context Gemini is an adjudication tool, not the primary architecture.
- Pattern names are clinical internally, neutral publicly. DARVO → "possible role-reversal pattern" in the UI. Ethics + legal.
- No training. AGON is inference-only. Corrections corpus accumulates; training is deliberate future work, gated on the corrections corpus reaching critical mass.
| Primitive | Definition |
|---|---|
| Actor | Any party capable of holding an interest or making a claim |
| Claim | An asserted fact, evaluation, or normative statement attributed to an actor |
| Interest | An underlying goal or need (Fisher/Ury distinction from "position") |
| Constraint | A rule, norm, or structural limit |
| Leverage | A resource, dependency, or capability that shifts bargaining power |
| Commitment | A promised future action, with subject and deadline |
| Event | A dated or orderable occurrence |
| Narrative | A coherent framing across multiple claims |
ASSERTED · DENIED · ACKNOWLEDGED · ACKNOWLEDGED_AMBIGUOUSLY · DENIES_SCOPE · COMMITS_TO · REVOKES · BLOCKS · ENABLES · CAUSES · PRECEDES · CONTRADICTS · SUPPORTS · CITES · HOLDS_INTEREST · FRAMES · LEVERAGES · CONSTRAINED_BY
Every edge carries a provenance field. Missing provenance fails validation.
When a predicted edge is close to but not identical to the gold edge, partial credit is awarded — ACKNOWLEDGED ↔ ACKNOWLEDGED_AMBIGUOUSLY = 0.75, BLOCKS ↔ CONSTRAINED_BY = 0.40, etc. See crates/tacitus-contracts/ for the full matrix.
AGON/
├── README.md ← you are here
├── docs/
│ ├── INDEX.md ← doc map (start here)
│ ├── AGON_GUIDE.md ← operator's guide (start/stop, costs, day-by-day)
│ ├── BUILD_PLAN_PERCEPTION.md ← 15-prompt build plan (~90 days)
│ ├── DEPLOYMENT_GCP.md ← target Cloud Run + Vertex topology
│ ├── EXTERNALS.md ← what you provide (Gemini-only)
│ ├── HONEST_STATE.md ← brutally honest accounting of what is real
│ ├── AUDIT_2026-05-13.md ← 15-finding code audit
│ └── INTEROP.md ← trinity integration (AGON ↔ DIALECTICA ↔ KAIROS)
├── PROJECT_LEDGER/
│ ├── AGON_LEDGER.md ← MVP v0.1.0 sprint (shipped)
│ ├── PERCEPTION_LEDGER.md ← 15-prompt perception sprint tracker
│ └── STATE.json ← current state, next prompt, open externals
├── crates/
│ ├── tacitus-contracts/ ← typed primitives + JSON Schemas (PROMPT 01) ✓
│ ├── aco-text/ ← canonical text + segmenter + quoted-speech FSM + speaker turns (PROMPT 02) ✓
│ ├── aco-time/ ← Allen-13 temporal algebra (PROMPT 03) ◐
│ ├── aco-lex/ ← hedge/modality/passive/pronoun extractors (PROMPT 04) ◐
│ ├── aco-encode/ ← BGE-M3 + DeBERTa-NLI + fastcoref (PROMPT 05) ◐ scaffolded
│ ├── aco-llm/ ← Vertex Gemini backend + retry middleware (live)
│ ├── aco-extract/ ← L1+L2+L3 perception pipeline (PROMPT 07) ☐
│ ├── aco-fuse/ ← cross-doc actor resolution (PROMPT 08) ☐
│ ├── aco-temporal/ ← commitment state machine (PROMPT 08) ☐
│ ├── aco-patterns/ ← DARVO ✓ · Anchoring ✓ · Conspicuous Absence ✓ · scope creep ☐ · coalition ☐ (PROMPT 09) ◐
│ ├── aco-score/ ← calibration + conformal prediction (PROMPT 10) ☐
│ ├── aco-prov/ ← lineage DAG + Merkle audit (PROMPT 11) ☐
│ ├── aco-storage/ ← Cloud SQL via sqlx (live)
│ ├── aco-server/ ← Axum + workbench UI (live)
│ ├── aco-cli/ ← agon-cli
│ ├── aco-core/ ← shared types + provenance
│ ├── aco-perceive/ ← MVP perception (refactored at PROMPT 07)
│ ├── aco-fuse/, aco-infer/, aco-embed/, aco-learn/, aco-bench/ ← MVP scaffold
├── infra/terraform/ ← VPC + Cloud SQL + Cloud Run + GCS + Eventarc + IAM
├── scripts/
│ ├── agon-up.sh / .ps1 ← turn ON
│ ├── agon-down.sh / .ps1 ← turn OFF
│ ├── agon-status.sh / .ps1
│ └── agon-nuke.sh ← terraform destroy
├── migrations/ ← Postgres schema
├── corpora/ ← test inputs
├── Cargo.toml ← workspace + deps
├── Dockerfile, compose.yaml
├── Makefile
└── .env.example
Legend: ✓ done · ◐ in flight · ☐ planned (see PROJECT_LEDGER/PERCEPTION_LEDGER.md)
| Phase | Days | Prompts | Deliverable |
|---|---|---|---|
| Foundations | 1–10 | 01–03 | Doc round-trip: normalize → segment → time extract → evidence spans verify |
| Encoders + LLM | 11–25 | 04–06 | Local ONNX (BGE-M3 / DeBERTa / fastcoref) + Vertex Gemini routing |
| Perception + patterns | 26–45 | 07–09 | Full pipeline emits ACO primitives + 5 named patterns with golden fixtures |
| Calibration + provenance | 46–60 | 10–11 | Calibrated confidence on every primitive · litigation-grade audit export |
| Prod deploy + UI | 61–75 | 12–13 | Split CPU/GPU services · corrections capture in workbench |
| Eval + adversarial | 76–90 | 14–15 | TCGC v0.2 + Inspect-AI + 80-case adversarial pack |
Hard sequencing: 01 → all · 02 → 03/04/05/07 · 05+06 → 07 · 07+08 → 09 · 11 → 14 → 15.
Full spec: docs/BUILD_PLAN_PERCEPTION.md (1246 lines, every prompt self-contained).
| If you want to… | Read |
|---|---|
| Full doc map | docs/INDEX.md |
| Copy-paste demo recipes (curl every endpoint) | docs/DEMO_RECIPES.md |
| See where AGON is going (standalone + trinity) | ROADMAP.md |
| Trinity integration (AGON ↔ DIALECTICA ↔ KAIROS) | docs/INTEROP.md |
| Run it for a few days then stop | docs/AGON_GUIDE.md §2 + §9 |
| Understand the architecture | docs/AGON_GUIDE.md §1 + docs/BUILD_PLAN_PERCEPTION.md |
| Know what AGON depends on externally | docs/EXTERNALS.md |
| Know what's deployed where | docs/DEPLOYMENT_GCP.md |
| See the typed primitive contracts | crates/tacitus-contracts/README.md |
| See what's done vs in-flight | PROJECT_LEDGER/PERCEPTION_LEDGER.md |
| Honest accounting of what's real | docs/HONEST_STATE.md |
| Component | Choice | License | Why |
|---|---|---|---|
| Embeddings | BGE-M3 | Apache-2.0 | Dense + sparse + ColBERT in one model · multilingual · ONNX-exportable |
| NLI | DeBERTa-v3-large-mnli (MoritzLaurer) | MIT | Best open NLI checkpoint · INT8 quantizable |
| Coreference | fastcoref | MIT | License-clean · 78.5 F1 |
| ONNX Runtime | ort 2.x |
MIT/Apache | Production-proven · pure Rust |
| Time extraction | hand-rolled Rust DFA | — | HeidelTime/SUTime are GPL — can't use |
| Segmenter | hand-rolled SRX-style | — | pragmatic-segmenter is MIT but we own the impl |
| Postgres | self-hosted on Cloud SQL | OSS | $25/mo at dev tier |
| Vector store | pgvector extension |
OSS | No managed vector DB |
| Annotation | Argilla | Apache-2.0 | Self-host on Cloud Run |
| Eval orchestrator | Inspect-AI (UK AISI) | Apache-2.0 | Principled |
| LLM observability | Langfuse self-hosted | MIT | Self-host vs $100/mo SaaS |
| Remote LLM | Vertex Gemini 2.5 Flash + Pro | paid | Schema-constrained, $0.30/$2.50 per M tokens |
Vendor strategy: Gemini-only. Cross-validation done with Flash vs Pro at different temperatures / prompt versions. Anthropic + OpenAI backends in the original plan were dropped 2026-05-13 (see docs/AUDIT_2026-05-13.md §F-10).
URL: https://agon-dev-tbryoen6qa-uc.a.run.app
User: AGON
Password: AGON
Status: https://agon-dev-tbryoen6qa-uc.a.run.app/api/info
Paste a multi-turn conflict (Slack thread, email reply chain, deposition snippet, board minutes). Click Perceive. Watch the friction matrix, the actor/claim graph, and the structured ACO primitives appear with verifiable evidence quotes.
BASE=https://agon-dev-tbryoen6qa-uc.a.run.app
AUTH="AGON:AGON"
# 1. Liveness — anyone can hit, no auth.
curl -s $BASE/healthz
# 2. Service info — version, deployment, db status.
curl -s -u $AUTH $BASE/api/info | jq
# 3. Full backend introspection — layers, ML strategy, registered patterns, doc index.
curl -s -u $AUTH $BASE/api/system | jq
# 4. Pattern detector catalog — names, kinds, descriptions, live vs planned.
curl -s -u $AUTH $BASE/api/patterns | jq '.patterns[] | {id, version, live, public_name}'
# 5. Pipeline map — 12 stages with crate + kind + p50 latency.
curl -s -u $AUTH $BASE/api/pipeline | jq '.stages[] | {order, id, crate, kind, p50_ms}'
# 6. Past perceptions.
curl -s -u $AUTH $BASE/api/sessions | jq '.sessions[0:3]'
# 7. Run a real perception.
curl -s -u $AUTH -X POST $BASE/api/perceive \
-H "Content-Type: application/json" \
-d '{
"text": "Sam (Mon): We agreed you own the Q4 deck by Thursday.\nAlex (Mon): Sounds good.\nAlex (Thu): I never said I would own it.\nSam (Thu): That is not what we discussed.\nAlex (Thu): You are putting words in my mouth.",
"title": "Q4 deck dispute"
}' | jq '{
elapsed_ms,
patterns: .patterns_detected[] | {pattern_id, public_name, raw_confidence, evidence_excerpts},
friction: .friction_matrix.pairs[0],
quality: .quality_gates
}'/api/perceive on the Q4 deck dispute returns (real, live):
{
"patterns_detected": [
{
"pattern_id": "darvo",
"pattern_version": "0.1.0",
"public_name": "possible role-reversal pattern",
"raw_confidence": 0.70,
"actors_involved": ["actor_alex"],
"evidence_excerpts": ["I never", "You're putting words in my mouth"],
"explanation": "Actor `actor_alex` denied at turn 2 (\"I never\") then attacked/reframed the accuser at turn 4 (\"You're putting words in my mouth\"). Classical role-reversal sequence."
}
],
"friction_matrix": {
"pairs": [{"a_label":"Alex","b_label":"Sam","heat":100,"reasons":["denial pressure","pattern: defensiveness","escalation signal", "..."]}]
},
"quality_gates": [
{"label":"Verified evidence coverage","status":"pass","detail":"33/33 primitive evidence quotes verified"},
{"label":"Actor ambiguity","status":"pass"},
{"label":"Conflict signal strength","status":"pass"}
]
}Open the URL in a browser. Login AGON / AGON. Paste a multi-turn dispute. Click Perceive. You'll see the same data rendered as a force-directed graph + friction matrix + named patterns + raw JSON inspector.
This is built in public by Giulio Catanzariti for TACITUS. The 15-prompt build plan is designed for Claude Code Opus 4.7 to execute one prompt per session, one PR each. If you want to participate:
- Pick the next prompt in
PROJECT_LEDGER/PERCEPTION_LEDGER.md - Branch
sprint/<NN>-<name> - Implement against the verification block in
docs/BUILD_PLAN_PERCEPTION.md - Open PR · the ledger row turns ✓ on merge
Issues with the spec? Open one tagged spec-drift.
MIT OR Apache-2.0, at your option. See LICENSE.
@software{agon2026,
author = {Catanzariti, Giulio},
title = {AGON: A Perception Engine for Human Conflict},
year = {2026},
url = {https://github.com/sargonxg/AGON},
note = {TACITUS},
}Maintainer: Giulio Catanzariti · giuliocatanzariti@gmail.com · TACITUS — making conflict legible.