Skip to content
MARS

Status — pre-release · in active development

You’ve arrived early.

MARS has not shipped as a product yet. What you’ll see on this site — the architecture, the runtime, the measured numbers — is real, but the product itself is still being built, gate by gate. We’re working on it actively, and we intend to put it in your hands soon.

In the spirit of this project: no promises without receipts. Everything described here is bound to published evidence — including what doesn’t work yet.

Track project status

Where it matters

Built for answers that carry consequences

Fluency is cheap. In commerce, operations and regulated work, a wrong answer is a refund, a breach, or a lawsuit. MARS is designed for exactly those rooms.

E-commerce

An agent that never invents a price

Chatbots that hallucinate prices and policies create real liability — courts have already held companies to what their bots promised.

The entire catalog lives in local Hyphae as versioned snapshots. The on-site agent either quotes the exact price from the current snapshot — with a receipt — or abstains and hands off to a human.

  • Price changes are snapshots — yesterday’s answer stays auditable.
  • Receipts double as a compliance audit trail per response.
  • Runs on your server. Catalog data never leaves the building.
  1. 01 Catalog + policies ingested as versioned sources
  2. 02 Question → deterministic intent → retrieval with proof
  3. 03 Exact span quoted, price bound to snapshot version
  4. 04 No evidence? Governed abstention → human handoff

Back-office agents

Small tasks, auditable actions, tiny hardware

Email triage, calendar coordination and routine replies don’t need frontier models — they need reliability, privacy and a paper trail.

Agents run on a small local server. Deterministic rules resolve the routine; the compact model drafts the rest; every action crosses the verifier. Draft-first: confident categories earn auto-send through gates.

  • Asynchronous by nature — a single small GPU serves a whole office.
  • Personal memory supports remember, correct and forget — with receipts.
  • Escalate the hard 5% to a frontier teacher, under the same verifier.
  1. 01 Company knowledge + preferences in local Hyphae
  2. 02 Trigger → rules first, model only when needed
  3. 03 Action proposal → verifier → policy check
  4. 04 Executed with receipt — or parked for a human

Regulated industries

Proof-bearing answers for rooms where proof is the product

Legal, medical, financial and defense contexts cannot accept “the model is usually right”. They need provenance, scope control, and the ability to run disconnected.

PRIVATE_LOCAL mode runs the full stack air-gapped: local inference, local Hyphae, offline-verified proofs. Visibility scopes enforce who may see what; expired or out-of-scope knowledge fails closed.

  • Air-gap friendly: no external inference, no external retrieval.
  • Forgetting is verifiable — deletion produces provable abstention.
  • Authority is deterministic code, not neural confidence.
  1. 01 Controlled corpus ingested with license + provenance
  2. 02 Tenant and visibility scopes enforced at retrieval
  3. 03 Answers bind to exact spans in authorized sources
  4. 04 Full receipt trail for every response — exportable

One architecture, many rooms

Every deployment is the same governed vertical: local model, verified memory, receipts end to end. Only the corpus changes.

Explore the technology