Fihan Systems · Trusted Autonomy Stack

One stack. Two trust boundaries. One ledger.

Fihan Edge and Interocept were built separately, to secure autonomy in two different places. Interocept already writes every gate decision to a hash-chained ledger. Put the two together, and a physical override could write into that same ledger as a tool-call gate does — a single, auditable chain of custody from what an agent is allowed to decide to what a machine is allowed to do. That second half doesn't exist yet; see below for exactly what's built and what isn't.

Two products
Built and sold independently today
Every gate call
Interocept logs it today, hash-chained and append-only
Self-hosted
Both run on infrastructure you own, start to finish
Half built
Interocept's ledger ships today; Fihan Edge's side of it doesn't yet
The gap today

Autonomy is certifiable, or it's capable. Rarely both.

A classical planner is provable but brittle. A learned policy adapts but resists certification. A software agent that can plan and act is powerful but ungoverned — one poisoned document, or one slow drift in its own goals, away from doing something no one signed off on. Assurance frameworks in regulated and defense-adjacent procurement are written for a stack that reasons about both halves of that problem together. Today, most vendors only cover one half.

The stack

Two layers, each already built to hold a line.

Neither product needs the other to work. Put together, they compose into one trust chain instead of two separate ones.

Interocept — agent layerFihan Edge — physical layerevidence ledger
Layer 1 — Agent

Interocept gates what an agent may command.

Before an AI agent's tool call reaches a real system, Interocept reads the model's own internal state and checks it against a deterministic policy gate. An agent trying to retask a sensor, escalate a response, or take an action it was never provisioned for is caught here — before the command exists anywhere else.

  • Reads internal activations, not the model's output text
  • Fail-closed by default on mutating, egress and destructive actions
  • Every check writes a hash-chained record — allow, escalate, or deny
Layer 2 — Machine

Fihan Edge enforces what the machine may physically do.

A deterministic Rust safety supervisor sits between the learned navigation policy and the actuators, with absolute veto authority. Whatever the agent above decided, the robot itself will not execute a move that violates a hard safety constraint.

  • An independent check on every single movement, with no override path
  • Sub-millisecond veto, measured in hardware-in-the-loop
  • Runs on-device — no cloud dependency, no backhaul required
Layer 3 — Record

One shared, hash-chained evidence ledger — today, only half of it.

Interocept already writes every gate decision — allow, escalate, deny — to a tamper-evident, hash-chained log. Fihan Edge's on-device safety supervisor enforces its physical veto today but doesn't log it to a ledger of its own; the closest existing implementation is Fihan Fleet's, a separate fleet-telemetry product, not Fihan Edge itself. Extending the same append-only pattern onto Fihan Edge's own overrides is the concrete piece of work between this architecture and a shipped joint ledger.

  • Interocept: every gate decision is already hash-chained and append-only
  • Fihan Edge: physical overrides are enforced but not yet logged to a ledger
  • Fihan Fleet (a separate product) already runs a working fleet-telemetry ledger — a nearer starting point than building one from scratch
Where it plays

Same two products, six mission profiles.

Each of these reuses existing capability — the extension is in framing and integration, not invention.

Led by Fihan Edge
CBRN and radiological detectionCounter-UAS false-alarm suppressionGPS-denied and subterranean navigationSensor chain of custody
Led by Interocept
SIGINT / HUMINT / OSINT copilotsAutonomous C2 and tasking gatesLong-duration ISR goal-drift watchAI accreditation and ATO evidence
Straight talk

What this page does not claim.

The same reason our product pages name their own limits applies here, maybe more so:

“A proven, fielded joint deployment.”

Fihan Edge and Interocept ship, are sold, and are deployed independently today. This page describes an architectural fit and a shared evidence-pack pattern — not a combined product that exists yet.

“The two products run on the same box.”

No integration work has been built between them — and it's not just wiring. Interocept already produces a hash-chained ledger; Fihan Edge's on-device safety supervisor doesn't log to one today. The nearest existing implementation is Fihan Fleet's ledger, a separate, separately-sold product, not Fihan Edge itself.

“This replaces evaluating either product on its own.”

Each product stands on its own technical merits and should be evaluated that way first. This page is for the case where you need both halves of the problem solved and want to know they'd compose cleanly.

Who it's for

Teams that need the whole chain, not one link.

  • Programmes that need one auditable trust chain from an agent's decision to a machine's action
  • Accreditors requiring layered, evidence-backed assurance for AI-directed autonomous systems
  • Operators already running a fielded autonomous system and the AI agents that task it
  • Anyone evaluating one product who wants to know how the other would fit alongside it

Which layer do you need first?

Tell us whether you're starting from the robot or the agent — we'll show you how the other half fits.