INTRODUCING JSOC AI — FABRIC GOVERNS EVERY AI ACTION. PROOFSTATE PROVES EVERY RELEASE.

EXPLORE JSOC AI
JSOC IT — Post-Security.
TRUST CENTER LET'S TALK

CHAPTER 04.1 — JSOC AI · FABRIC

FABRIC

THE AI CONTROL PLANE FOR REGULATED ORGANIZATIONS.

REQUEST A WALKTHROUGH FOLLOW ONE REQUEST
FABRIC · GATEWAY · POLICY DECISIONSLIVE
Four actions reach the Fabric gateway: one is allowed, one waits for a person to approve, one has a card number masked, and one is blocked. Every decision is written to the audit trail.AUTOMATION STEPSPOLICY GATECONNECTIONSSLACKPAYDESKCRMUNKNOWNslack.post_messagepaydesk.issue_refundcard 4111 1111 1111 1111card •••• •••• •••• 1111http.post unknown.io✓ ALLOW? ASK A PERSON▒ MASK× BLOCKWAITING…APPROVED ✓AUDIT TRAIL → PROOFSTATE
01WHAT IT IS
A gateway and automation platform for AI: governed agents, workflows and data sources behind one door.
02WHO IT IS FOR
Regulated organizations that want AI doing real work — with security, finance and compliance comfortable about it.
03WHAT YOU GET
Rules on every action, a person in the loop when it matters, spend under control, and a record of everything.

/ THE REAL PRODUCT

SEE IT
RUNNING.

Real screens · demo data · open any of them

/ THE PROBLEM IT SOLVES

AI WIRED STRAIGHT IN
IS A RISK.

Flip the switch to see the difference one governed gate makes.

FIG. — HOW AI REACHES YOUR SYSTEMS
AUDIT TRAIL → PROOFSTATE · 0 DECISIONS RECORDED! KEYS IN CODE! NO LIMITS! NO RECORD! DATA LEAKAI AGENTFABRICGITHUBJIRAPOSTGRESQLZENDESKSTRIPESNOWFLAKEUNKNOWN.IOALLOWALLOWMASKALLOWASKMASK
EVERY AGENT HOLDS ITS OWN KEYS. NOTHING STOPS A LEAK, AND NOBODY CAN SAY WHAT IT DID.ONE GATE. A VERDICT ON EVERY ACTION, A PERSON WHEN IT MATTERS, THE LEAK BLOCKED, AND A RECORD OF EVERYTHING.

/ FOLLOW ONE REQUEST

ONE COMPLAINT.
FIVE STEPS. ALL GOVERNED.

Scroll through a real-world example: a customer reports a double charge, and an AI-powered automation handles it.

  1. 01

    / SOMETHING HAPPENS

    A customer writes in: “I was charged twice.”

    The shop’s contact form posts to Fabric, and the “Customer complaints” automation starts a run. Every run names the exact version of the automation it ran.

    • START · WEBHOOK
    • AUTOMATION v7
  2. 02

    / AI INVESTIGATES — WITHIN LIMITS

    An agent looks into it, using only what it was given.

    It can read orders and track the parcel because those actions were granted. The card number on file is masked before any model sees it.

    • orders.read → ALLOW
    • parcel.track → ALLOW
    • card_number → MASK
  3. 03

    / THE MODEL DRAFTS A REPLY

    Using the reply prompt kept in ProofState.

    Not whatever was last pasted into a script: the version labelled “production”, with its test cases and judges behind it.

    • PROMPT support-reply
    • LABEL production
  4. 04

    / A PERSON APPROVES THE MONEY

    Refunds over $50 need a manager.

    A policy says so, so Fabric pauses the run and asks in Slack. One click approves it, and who approved it is recorded with the run.

    • paydesk.issue_refund $84
    • ASK A PERSON → APPROVED
  5. 05

    / EVERYTHING IS PROVEN

    The reply goes out. The evidence stays.

    The run lands in the audit trail and in ProofState, where judges score the reply for policy and tone. Next week’s better prompt starts here.

    • AUDIT TRAIL
    • PROOFSTATE · JUDGED

/ POLICIES

FOUR ANSWERS.
EVERY SINGLE TIME.

Every action and every model call gets one of four answers before it runs.

  • ALLOW

    Within the rules. It runs, and the decision is recorded.

    EXAMPLEPost the daily summary to #support.

  • ASK A PERSON

    Allowed with a human in the loop. The run pauses until someone approves.

    EXAMPLEIssue a refund over $50.

  • MASK

    Allowed, but sensitive fields are hidden before any model sees them.

    EXAMPLERead a customer record with a card number.

  • BLOCK

    Not allowed. It never happens, and the attempt is recorded.

    EXAMPLESend data to an unknown external host.

/ BUDGETS

SPEND WITH
A CEILING.

Set a monthly budget for an organization, a person or a single automation, and decide what happens as it fills up.

  1. 01
    WARN

    The owner hears about it early.

  2. 02
    DEGRADE

    Steps move to a cheaper routing group automatically.

  3. 03
    PARK OR BLOCK

    Runs wait for approval, or stop — your call.

BUDGET · AUTOMATION "MORNING NOTE"MONTHLY CEILING
WARNDEGRADE → fabric-bulkPARK · BLOCKSPEND

/ THE BUILDING BLOCKS

FABRIC IN
PLAIN ENGLISH.

Eight words, and you know how Fabric thinks.

AUTOMATION
A job Fabric does for you: how it starts, the steps it runs and the rules it runs under. One on/off switch.
STEP
One thing an automation does, in order: read a database, ask the model, call an API, wait for a person.
AGENT STEP
A step that pursues a goal by reasoning — but only with the actions it was given, under the same policies.
CONNECTION
A system Fabric reaches on your behalf: an app or API, an MCP server, or a database.
ACTION
One thing a connection offers, with a clear name like slack.post_message. Policies are written against these.
POLICY
A rule over actions and model use: allow, ask a person, mask, or block.
BUDGET
A monthly spend ceiling, and what happens when it is reached.
RUN
One execution from start to result. The thing you watch, approve, pay for — and prove.

/ CONNECTIONS

EVERYTHING PASSES
THROUGH ONE GATE.

Apps, APIs, MCP servers and databases — each connection exposes named actions, and every action meets your policies first.

Examples from Fabric's connection catalog. Brand names belong to their owners.

YOUR OWN DEPLOYMENT

Each client runs its own Fabric — never a shared tenant. Organizations inside it keep departments, subsidiaries or environments apart.

ANY MODEL, ONE GATEWAY

Steps name a routing group, not a vendor. Change providers or fall back to a cheaper model without rewriting automations.

EVIDENCE BUILT IN

Every run is sent to ProofState, where it is traced and judged. Prompts live there too, versioned and labelled.

SEE PROOFSTATE

/ FREQUENTLY ASKED QUESTIONS

QUESTIONS
ABOUT FABRIC.

IS FABRIC AN AI MODEL?

No. Fabric is the control plane around models. It decides which model a step uses, what that step may touch, what it may spend, and when a person must approve — and it records every decision.

WHERE DOES FABRIC RUN?

Each client gets its own Fabric deployment, never a shared tenant. Inside it, organizations separate departments, subsidiaries or environments, each with their own policies and budgets.

CAN OUR EXISTING TOOLS CALL FABRIC?

Yes. Other software can call into Fabric with an API key — for example the OpenAI SDK or Claude Desktop — and still be governed by the same policies and budgets.

WHAT DOES FABRIC SEND TO PROOFSTATE?

Every run: its steps, model calls, policy decisions and results. Prompts can live in ProofState too, versioned and labelled, so a step always uses the production version.

YOUR NEXT STEP

PUT ONE AUTOMATION BEHIND FABRIC.

Pick a workflow your team would love to hand to AI. We'll walk through the policies, approvals and budget it would run under.

REQUEST A FABRIC WALKTHROUGH BOOK A CONVERSATION