Reference architecture · managed early access

One operational layer for the physical world.

A building panel here. An order book there. Sensors nobody reads until it is too late. Fibric defines a governed pattern for reading that state, proposing a response, checking policy, and dispatching only through approved integrations. BearScope is the live production proof.

Get started → See the loop

One control contract. A chiller, an open order, a door sensor, a support queue.

The control loop

Sense, reason, act. A loop designed to close under governed conditions.

01 / SENSE

Unify supported inputs into one picture

Your operation already produces the data. It just lives in systems that do not always agree. The reference architecture normalizes governed source events so an operator can reason over a bounded operational picture.

Reads software systems and physical devices through the same connector model.
Normalizes supported connector events into a canonical envelope with explicit source and tenant context.
Source-tagged measurements, with fallback and missing states kept visibly separate.

02 / REASON

Understand, predict, and plan

A base model reads the governed context and proposes what may be happening and what to do next. The proposal is structured into explicit steps so policy and a person can review it before any enabled action runs.

Predicts needs and risks instead of waiting for an alarm.
Plans across systems. One intent can touch HVAC, an order, and a team.
In the reference path, model output becomes a structured proposal rather than holding direct dispatch authority.

03 / ACT

Take the action, keep the record

A deterministic path checks the plan, then dispatches only the approved capabilities exposed by connected systems. Single-flight and idempotency controls reduce duplicate effects, and the reported result is kept with the proposal.

Dispatches only through validated capabilities exposed by supported connected systems.
Single-flight and idempotency controls reduce duplicate and runaway-action risk.
Supported executed steps can leave an attributable record of what the Fibric boundary observed and reported.

The loop, closing

1event shape

Sense

The reference contract gives supported source events a canonical envelope; each connector still requires validation.

1in flight

Single-flight

One action per entity in flight, with idempotency keys to suppress known retries.

Connectorfirst

Act

Compatible existing systems may be reused; a deployment can still require gateways, upgrades, or independent safeguards.

Supportedpaths

Audit

Instrumented execution paths preserve the proposal, policy result, and connected system's reported outcome.

Operators

Named AI workers that run one operation, end to end.

An operator is a named role with a bounded job. It watches approved systems, proposes what their state may mean, and uses only the capabilities enabled for that deployment. BearScope runs two operator roles today; the rest of the catalog is managed early access.

Live today · BearScope

Order Risk

a paper & print company

BearScope combines supported order, conversation, and production signals to score and surface open orders that may miss their ship date upstream of the carrier scan. Where a managed action is separately configured and authorized, the path may propose a hold, customer nudge, or owner escalation; availability depends on connector capability and policy.

Senses: supported order · proof · production signalsReasons: which order may slip, and whyManaged option: propose hold · nudge · escalate

OPR

Comfort & Energya hotel

Illustrative reference pattern. A managed deployment could combine validated occupancy, weather, and building-system inputs to propose bounded comfort or energy changes. Building, access, and life-safety controls remain subject to independent safeguards and site acceptance.

Potential inputs: BMS · occupancy · weatherPotential managed actions: setpoints · lighting

OPR

Floor Flowa warehouse

Illustrative reference pattern. A managed path could combine supported station, lane, exception, and inbound signals, then propose rerouting or scheduling changes for human or policy review. Real capabilities and recovery behavior require WMS validation.

Potential inputs: WMS · stations · inboundPotential actions: reroute · reschedule · notify

One live operator and two reference patterns show how the same control contract can be adapted across industries.

Open the marketplace →

The same loop, any floor

The operator does not know it is a warehouse.

This warehouse sequence is a reference pattern: read stations and inbound, identify where flow may break, and propose a lane change. A real deployment still requires governed connectors, policy, and site-specific acceptance testing.

✓ Enabled actions are intended to clear policy before dispatch

Connectors & capabilities

Plug into the real world. Swapping a system is config.

Ticketingcapability

→

KustomerZendesk

Commercecapability

→

MagentoShopify

Building controlcapability

→

BACnetNiagara

Voice & contactcapability

→

Amazon Connect

In the reference model, an operator asks for a capability such as “ticketing” rather than naming Kustomer. A Kustomer-to-Zendesk change still requires capability parity, authentication, data mapping, policy review, migration planning, and acceptance testing; affected operations may pause until those checks pass.

Capability over connector

In the reference model, an operator asks for a capability such as ticketing or building control rather than naming a vendor. Swapping systems is intended to be a connector configuration change, subject to capability parity and validation.

Built on MCP

The connector contract is designed around MCP-shaped integrations for software and physical-system interfaces. The public SDK and generalized runtime are not yet available; connectors are provisioned with the Fibric team in managed early access.

Connector-first, not hardware-free

A managed deployment starts with supported systems already in place. Depending on protocol, network, safety, and control requirements, it may still need a gateway, upgrade, independent safeguard, or other site work before an action is enabled.

Browse integrations →

The reference engine

The AI proposes. A deterministic check disposes.

This is the contract the generalized kernel is meant to enforce. Today BearScope implements the governed pattern directly in its production BFF; the reference kernel is not yet the live request path.

The AI

Proposes

Reads the live picture and returns a validated ExecutionPlan: a sequence of steps, each one explicit. It suggests; it does not decide.

policy gate→fail-closed

The check

Approves or blocks

Evaluates supported plan steps against configured policies, then dispatches only enabled steps that pass. A covered rule can block the capability within the validated Fibric boundary; downstream systems retain their own behavior.

Fail-closed policy

A configured rule can veto an enabled step

For a validated managed path, a proposed step that fails a covered rule should be withheld before dispatch or sent for review. The deployment must test that fail-closed behavior and define limits outside the Fibric boundary.

Single-flight + idempotency

Duplicate effects are actively suppressed

One action per entity is allowed in flight and each action carries an idempotency key. That collapses known retries and limits concurrency; downstream-system behavior still matters.

Walled-off tenancy

Tenant scope is enforced in data access

The live BearScope path sets tenant context before governed queries, and row-level security protects tenant-owned tables at the database layer.

Full audit trail

Supported actions keep their evidence

Where instrumented, supported executor steps preserve what the Fibric boundary observed, proposed, approved, and received as the connected system's reported outcome. That record supports review without proving every downstream fact.

Read the trust spine →

Studio is the product direction.

For the people who run the place

Studio is the planned builder experience for connecting a system and describing an intent in plain language. It is not a public self-service product today; early deployments are configured and reviewed with a Fibric engineer.

Scope a governed connector and its permissions with the deployment team.
Describe the outcome you want, not the steps to get there.
See the proposed plan, approve it, and keep the record.
Configure policies for each supported operator path and validate their enforcement before enabling actions.

Request a managed pilot →

Studio · new intent

Hold any order that won't ship on time, and tell the owner why.

1Read open orders and production statussense

2Score each order's risk of missing its datereason

3Place a hold on at-risk ordersact · vetted

4Notify the account owner with the reasonact · vetted

Receipt #4471 · 4 steps · 0 vetoed Approve & run

In production, on real orders.

BearScope · a paper & print company

It catches the order before the carrier scan ever does.

BearScope is the live production proof: CX operations across conversations, calls, QA, Radar, and connected order-risk evidence. Its shipping BFF implements the governed pattern directly today; the generalized kernel remains the reference path being built toward.

.82

SO-48117 Mercer Dental

Carrier not scanned 26h past pickup · ships tomorrow

Hold + notify

.61

SO-48140 Halcyon Press

Stock short 1,200 of 2,000 on hand

Split ship

✓ Governed source data · policy check · record on supported dispatch

See BearScope →

Managed early access

Bring reason to what you run.

Tell us what you operate. We onboard the operation with you, end to end.

Managed deployments define what is checked, recorded, and recoverable before any action is enabled.