ACCOUNTABILITY LAYER FOR FEDERAL AI · DETERMINISTIC · SCIF & AIR-GAP READY

The decision engine for the contested edge.

Forge is the independent accountability layer for federal AI. Every agent action and autonomous call clears it, and when the operator cannot be in the loop, it makes the call on the asset and signs the record.

The same deterministic engine that holds AI agents accountable in the commercial world is the accountability layer for AI on the federal side. Point any federal AI agent or autonomous system at the Forge API and every high-consequence action clears the same independent checkpoint. Given the evidence and the rules of engagement, it makes the operational call (proceed, divert, abort, or maneuver), dictates the next action, and emits a signed, offline-verifiable record of why. No generative model in the decision path. No external calls. It runs in a SCIF, on an air-gapped network, or at the edge.

Proceed · divert · abort · maneuver On the asset, at the edge SCIF · air-gapped · offline-verifiable Deterministic · no generative model
DECISION PATHDETERMINISTIC RECORDSED25519-SIGNED DEPLOYMENTSCIF / AIR-GAP GENERATIVE MODEL IN PATHNONE
Why it matters

Autonomy without accountability is a liability.

Contested and denied environments are pushing the decision onto the platform, where no human can react in time. The hard part is not making a call. It is being able to prove, after the fact, that the call followed the rules of engagement, to a commander, an IG, or a coalition partner. A model's own logs cannot do that. Forge is the independent, deterministic record that makes an autonomous decision defensible.

It decides when no one else can.
And it can always show why.

Federal posture

Evaluated where it counts.

Forge has been put in front of the technical authorities who scrutinize mission decision systems, and is positioned to be bought.

/01

NRO

Independently evaluated by the National Reconnaissance Office on a bounded cislunar space-domain-awareness use case.

/02

JHU APL

Technical evaluation with the Johns Hopkins Applied Physics Laboratory.

/03

DIU

Engaged in the Defense Innovation Unit AI portfolio.

/04

CDAO Tradewinds

Assessed post-competition, readily awardable on the CDAO Tradewinds Solutions Marketplace.

/05

SAM.gov

Registered and active. CAGE code 1ASL5. U.S.-built, founder-owned.

/06

MBRSC

Mutual NDA executed with the Mohammed Bin Rashid Space Centre for an allied lunar evaluation.

Where it runs

One engine, across the mission set.

The decision layer is domain-agnostic. It sits above the GN&C, the autonomy stack, and the sensor fusion. It never replaces flight control or interlocks, and it signs every call.

/01

Lunar & planetary surface robotics

The decision and autonomy layer for surface operations: when to relocate, sample, divert, or hold, on a robot that cannot wait for a ground loop. In development against real mission profiles.

/02

Satellites & space domain awareness

Sensor-fusion and tasking decisions for spacecraft and SDA: weigh the track, make the maneuver-or-hold call, and sign the rationale for the operations center.

/03

Domain & mission systems

A signed decision layer above mission software and C2 (proceed, escalate, or reject), with the evidence chain attached, for air, maritime, subsea, and ground.

/04

PNT & resilience

Decisions under degraded or spoofed timing and navigation: choose the source, flag the anomaly, and record what was trusted and why.

/05

Contested & denied logistics

When the link is gone, the engine makes the bounded call locally and reconciles a signed record when connectivity returns.

/06

AI-agent oversight, in the SCIF

The same agent-accountability checkpoint from the commercial side, for autonomous and AI-enabled systems inside the perimeter.

/07

Synthetic and simulation training data

A signed bill of materials for simulation-generated and synthetic training data: how it was produced, where it was benchmarked, what remains uncertain, and the named next tests. A program office or a test and evaluation authority verifies it offline.

Responsible AI

Built around Traceable and Governable.

Forge maps directly to the DoD Responsible AI tenets (Responsible, Equitable, Traceable, Reliable, Governable). It gives a program the artifact those tenets ask for, the kind of evidence the CDAO RAI Toolkit and NIST SP 800-53 controls expect a system to produce. Forge does not claim a certification or an ATO. It produces the signed proof that supports one.

/01

Traceable

Every decision carries its evidence, the rules applied, the uncertainty that remained, and a signed, replayable record. Personnel can understand how the call was reached and reconstruct it later, offline.

/02

Governable

The human-review boundary is encoded in the policy, not left to the model. When a call exceeds the bound, Forge holds or escalates, and the record shows exactly where a human was, or had to be, in the loop.

/03

Reliable

Deterministic Bayesian logic with no generative model in the signed path. The same canonical input reproduces the same decision and the same replay key, so behavior is testable and the record does not drift.

Deployment

Built to run where the cloud cannot.

No generative model in the decision path means nothing to phone home and nothing to hallucinate. Forge is deterministic, self-contained, and verifiable with no access to us.

/01

SCIF & air-gapped

Runs fully offline on classified or isolated networks. No outbound calls, no external dependency at decision time.

/02

On the edge / on the asset

Small, deterministic footprint that runs on the platform, not in a data center, so the call survives a denied link.

/03

Offline replay

The same canonical inputs re-derive the same decision and replay key. An auditor can verify the signed record offline without asking Forge to vouch for it later.

/04

U.S.-built & ITAR-aware

U.S.-built and founder-owned, with export-control handled deliberately, not as an afterthought.

One core

The same engine. Commercial and federal.

Forge is one deterministic decision-and-record core. The vocabulary changes; the engine does not. That is the proof it is genuinely agnostic: the same diligence-grade record comes out no matter what goes in.

/ COMMERCIAL

Accountability for AI agents & high-consequence workflows

Agents · finance · insurance · critical infrastructure

The independent checkpoint every agent action and high-consequence decision clears. Forge decides, then signs the record.

/ THE ENGINE

Deterministic decision & signed record

No LLM in the signed path · offline-verifiable

Evidence and rules in. The call, the reasoning, and a signed, replayable record out. The same core, every domain.

/ FEDERAL

The operational call at the contested edge

Space · lunar · PNT · domain systems · SCIF

Makes the proceed / divert / abort / maneuver call on the asset when the operator cannot, and signs why.

How to engage

Built to be bought, not just briefed.

Forge is already evaluated, registered, and awardable. There are clean paths to put it under contract, and a small first step that proves it on your own ground before anything bigger.

/01

CDAO Tradewinds

Assessed post-competition and readily awardable on the CDAO Tradewinds Solutions Marketplace. A fast, competed path to award, never a sole-source claim.

/02

OTAs, consortia, and SBIR

Reachable through Other Transaction Authority vehicles, consortium memberships, and SBIR or STTR. We can plug into the path your program already uses.

/03

A bounded first step

One workflow, one decision, one signed record, one packet your reviewers can verify, on synthetic or sanitized data. Proof before scale, on a fixed scope.

/04

Direct briefing

A technical walkthrough for the program office, the technical authority, and the contracting shop together, tuned to your mission and your environment.

SAM.gov active / UEI FNMHTCGY8FK3 / CAGE 1ASL5 / U.S.-built, founder-owned

NAICS and PSC on request. Readily awardable, never sole-source.

Federal FAQ

The questions a program office asks.

Straight answers on what the accountability layer for federal AI is, how it maps to Responsible AI, where it runs, and who has evaluated it.

F.01What is an accountability layer for federal AI?
It is one independent, deterministic checkpoint that every federal AI agent action and autonomous call clears before it stands. Forge evaluates the action against the mission rules and the rules of engagement, makes the operational call when the operator cannot, and signs a record of the decision and the reasoning. Because there is no generative model in the signed path, a commander, an inspector general, or a coalition partner can replay the inputs and reproduce the exact record offline. That independent, reproducible record is the accountability the model's own logs cannot provide.
F.02How does Forge map to the DoD Responsible AI tenets?
Forge produces the artifact the Traceable and Governable tenets ask for. Every decision carries its evidence, the rules applied, the uncertainty that remained, and a signed, replayable record, so a person can understand and reconstruct how the call was reached. The human-review boundary is encoded in the policy, so the record shows exactly where a human was, or had to be, in the loop. That is the kind of evidence the CDAO RAI Toolkit and NIST SP 800-53 controls expect a system to produce. Forge does not claim a certification or an ATO. It produces the signed proof that supports one.
F.03Can Forge run in a SCIF or on an air-gapped network?
Yes. Forge is deterministic and self-contained, with no generative model in the decision path and nothing to phone home. It runs fully offline on classified or isolated networks, on the edge, or on the asset, with no outbound calls and no external dependency at decision time. Records verify with standard Ed25519 signature checks and replay without any network access to Forge, which is exactly what a SCIF, on-prem, or air-gapped environment requires.
F.04What does Forge do when the operator cannot be in the loop?
In a contested or denied environment the decision falls onto the platform, where no human can react in time. Given the evidence and the rules of engagement, Forge makes the bounded operational call locally (proceed, divert, abort, or maneuver), dictates the next action, and signs why. It never replaces flight control, the autonomy stack, or the safety interlocks. It sits above them as the decision and record layer, and when connectivity returns it reconciles a signed record an operations center can verify.
F.05Which authorities have evaluated Forge?
Forge was independently evaluated by the National Reconnaissance Office on a bounded cislunar space-domain-awareness use case, went through a technical evaluation with the Johns Hopkins Applied Physics Laboratory, and is engaged in the Defense Innovation Unit AI portfolio. It was assessed post-competition and readily awardable on the CDAO Tradewinds Solutions Marketplace, which is a competed path to award and never a sole-source claim. Forge is registered and active on SAM.gov under CAGE code 1ASL5, and is U.S.-built and founder-owned.