CLASSIFICATION-AWARE · AIR-GAPPED · CITED

Sovereign wargame
for allied ministries.

Pythia is a defense-AI platform that refuses to fabricate. Production game theory, framework-enforced citation discipline, and sovereign model substitution — running on your hardware, behind your wire.

Request a briefing See the platform →

Solvers in production: 6
Sovereign model lines: Gemma · Mistral
Citation floor: Framework-enforced
Deployment: Air-gapped

§ 01 — The problem

The next defense-AI scandal is a hallucination waiting to happen.

Every defense AI today is built on a brittle assumption: that confident language-model output is analysis. It isn't. The pattern is now visible across every major platform — models are allowed to invent, then their inventions are dressed in dashboards and presented as intelligence.

The cost is asymmetric. A wrong claim in a slide is a learning moment. A wrong claim in a kill chain is a war crime. A wrong claim in an Article-36 review ends the program.

Allied ministries feel it first. They need first-tier reasoning capability without the opacity, vendor lock-in, or customer-concentration risk that comes with it. And nobody else ships the math that would fix it.

§ 02 — Pythia

Refuses to fabricate. Mathematically.

Pythia is built around three refusals. Each one is enforced in code, not in guidance. Each one is a property of the platform that survives every prompt, every analyst, every deployment.

Cited everything

No claim ships without a click-through source. If retrieval returns nothing, Pythia says "no source" — and stops. Framework-level refusal, not prompt engineering.

Rational adversary by design

Every Red is modeled by Nash, QRE, or Bayesian-typing solvers seeded with cited doctrine. No black-box adversary. No defense AI ships production game theory today. We do.

Sovereign by construction

Gemma 3 for US / FVEY. Mistral for EU. Air-gapped on customer hardware. No training data, inference data, or telemetry leaves the customer boundary. Ever.

§ 03 — The product

One platform. Two inseparable halves.

A · Cited Synthesis

Ask anything. Get a sourced answer.

Analysts query in natural language across a multi-source corpus — OSINT, SOCMINT, partner-CERT feeds, treaties, AARs, drone footage, intercepts. Pythia returns a cited answer in seconds. Multi-lingual end-to-end: an English query can pull Mandarin, Russian, or Persian source.

Cross-modal: text, image, audio, video, tables
Snapshot hashes survive re-indexing
Coalition-aware redaction at retrieval time

B · Rational Wargame

Three COAs. Equilibrium math. Honest labels.

Operators pick a mission archetype and a Red doctrine. Pythia returns three Course-of-Action candidates with formal equilibrium payoff analysis, the Red-Team agent's worst-case counter-strategy, and a counterfactual ledger of alternatives.

Six solvers: Nash · QRE · Bayesian typing · IBR · replicator dynamics · backward induction
Doctrine-typed adversaries with citation pedigree
Living re-simulation as new intel arrives

Inseparable in practice. Wargame doctrines are grounded in the cited corpus. Recommendations are stress-tested by the Red-Team agent. Neither half alone is the product.

§ 04 — Architecture

Four subsystems. One coherent surface.

Pythia-Foundry

The action ontology

Nouns (entities) · Verbs (actions) · Policies (ROE) · Provenance · Classification · Counterfactual ledger. Every fact and every action is typed, audited, and replay-able.

Pythia-Wargame

Production game theory

Six solvers: Nash, QRE, Bayesian typing, IBR, replicator dynamics, backward induction. Doctrine catalog. Red-Team agent. Living re-simulation on new intel.

Pythia-KB

Tier-zero retrieval

Multi-modal agentic retrieval as an external service. Framework-enforced citation floor. Cross-modal across text, image, audio, video, and tabular sources.

∾

Pythia-Watch

The analyst surface

NL query · COA comparison · auto-SITREP · JAG quick-look · coalition redaction · multi-channel liaison across Slack, Teams, and LINE.

§ 05 — Auditable by construction

Article-36 ready. Replay-able by design.

AAR, Article-36 review, ICJ-equivalent proceedings — every one of them gets a single, signed, replay-able answer.

Provenance chain

Every claim is sourced to corpus chunks at retrieval time. Snapshot hashes survive re-indexing. Every assertion in every SITREP is click-through to its source.

Doctrine pedigree

Every Red adversary identifies which doctrine bundles seeded it, with citations to open-source manuals or partner-shared analyses. No black-box adversary models.

Honest solver labels

IBR is never labeled Nash. Replicator dynamics is never labeled equilibrium. The math is labeled by what it computes — never inflated into claims it cannot make.

Counterfactual ledger

Every chosen COA is logged with its alternatives and a predicted-vs-actual delta. Every decision survives review because the alternatives survive with it.

§ 06 — Who we serve

Where the incumbents cannot go.

Allied ministries that want first-tier reasoning capability without export-restriction friction, data-sovereignty exposure, or customer-concentration risk.

FVEY partners

Allied ministries across the Five Eyes coalition operating under data-sovereignty constraints that exclude customer-concentrated US platforms.

NATO non-FVEY

European defense ministries riding the ReArm Europe procurement window with a sovereign-AI imperative and Mistral-aligned model lineage.

JADC2-adjacent Asia

Indo-Pacific partners that need multi-language sovereign deploy — Mandarin, Korean, Japanese — and rapid integration with mixed US/local stacks.

§ 07 — Team

Built by people who have shipped together before.

Yale Lee

Co-Founder & CEO

Full-stack AI, security, and mission-critical systems leadership. Previously at Cisco and Lilee Systems.

Jeffrey Chang

Co-Founder & COO

Systems integration, operations, customer execution, and mission-critical fleet-dispatching domain expertise. Previously at Cisco and Lilee Systems.

Proven shipping cadence. Co-founding team with a track record of building together across two prior companies.

Domain insiders. We have lived our customers' workflow, security posture, and review-board pain.

Execution velocity. Platform shipped. First paid pilot scoped with a premier allied defense institute.

§ 08 — Engage

Brief us. Or let us brief you.

We work with allied ministries, defense institutes, and coalition staffs across FVEY, NATO, and the Indo-Pacific. If your mandate touches J2, J5, or J7 — and the words auditable, sovereign, and citation are non-negotiable — we should talk.

Pythia@Nurol.AI