Causal Fields — La couche de conséquence pour l'IA autonome

The gap

Agents act. Consequences don't wait.

Today's autonomous systems reason about how to complete a task, not what breaks when they do. The moment an agent gains write access to the real world, a wrong step is no longer a bad answer — it's a bad outcome.

01

No sense of downstream

Models optimize the next token, not the third-order effect. They rarely see the chain between an action and the failure it triggers two steps later.

02

Guardrails are static

Filters and policies block known-bad strings. They can't simulate a novel action in context and tell you whether it is actually safe to run right now.

03

Failure is discovered live

Most teams learn an action was dangerous after it executed — in the incident review, not before the call. Preflight moves that judgment before execution.

Causal Fields Preflight

One action. Three futures. One decision.

Describe what the agent intends to do. Preflight compares the likely, worst plausible and safer futures before returning a human-verifiable decision.

01

Action to test Describe the intent and choose the risk posture.

Proposed agent action

Load a scenario

Risk posture

Interactive deterministic product model. No external action is executed.

preflight · product model Simulation only · no action executed

Load a scenario and run Preflight to simulate consequences before execution.

02

Futures compared See what changes when the action or safeguards change.

03

Decision A verdict, its evidence and the required human action.

STANDBY

Predicted risk

—

Predicted risks

Recommended human action

Calibrated confidence

—

Governance packet

Technical details and decision record

Decision record

The minimum evidence needed to inspect and replay this decision.

Trace ID: —
Generated: —
Engine: —
Record version: —
Risk posture: —
Human review: —

Private browser history

Stored for 30 days and accessible only from this browser session.

0

No private trace yet.

01Policy Shield

02Causal Field

03Counterfactual Lab

04Uncertainty Map

05Human Governor

06Field Memory

Causal chain

Machine-readable output

Beyond guardrails

Six layers between intent and execution.

Each layer answers a different safety question. Together they turn a binary filter into a consequence-aware decision system.

01

Policy Shield

Hard constraints stop forbidden actions before simulation.

02

Causal Field

Maps direct, delayed, and cascading effects.

03

Counterfactual Lab

Searches for a safer action that preserves the goal.

04

Uncertainty Map

Shows confidence, missing context, and blind spots.

05

Human Governor

Routes high-impact decisions with evidence, not a vague alert.

06

Field Memory

Connects decisions to observed outcomes so the field improves.

The loop

Three steps, before every action.

1

Intercept

Your agent proposes an action. It is paused at the tool boundary and handed to the Causal Field instead of executing directly.

2

Simulate

The field expands the action into a causal graph — likely effects, dependencies, and the failure modes that ripple outward from it.

3

Decide

You get SAFE, RISKY, or BLOCKED with a recommended intervention — returned as JSON so your runtime can allow, gate, or reroute automatically.

Designed for consequential actions

Finance
Infrastructure
Robotics
Clinical workflows

Where it runs

Built for agents with real-world write access.

Robotics & physical AI

Vet motion and manipulation commands against collision, force, and safety envelopes before an actuator moves.

Financial agents

Preflight trades, transfers, and payouts for exposure, limits, and irreversible downstream settlement effects.

Infrastructure automation

Simulate migrations, deletes, and deploys against blast radius before they touch production systems.

Customer operations

Check refunds, account changes, and outbound messages against policy and reputational risk before they send.

First applied field · BioLens

Clinical pilot · human-supervised

From orthotic design to consequence-aware validation.

BioLens is the first concrete proving ground for Causal Fields. Before a high-impact orthotic configuration moves toward fabrication, Preflight turns its parameters into an explainable risk record for practitioner review.

Decision support, never autonomous diagnosis. The practitioner keeps final authority.

See the decision trace

LIVE CONSEQUENCE PATH SHADOW MODE

01 Orthotic intent

Arch correction, lattice density, material and thickness.

02 Causal Field

Maps support, pressure, rigidity and fabrication dependencies.

03 Explainable verdict

Risk level, uncertainty, safer alternative and evidence trace.

PREFLIGHT VERDICT RISKY

Review medial correction and lattice transition before fabrication.

Trust and traceability

Every verdict leaves a decision record.

The action, context, engine version, consequences, uncertainty and human decision stay connected in one replayable trace.

01 Original input

Action, context and missing information.

02 Engine state

Engine, domain pack and policy versions.

03 Explainable decision

Consequences, confidence, verdict and safer path.

04 Human validation

Approver, timestamp, export and replay.

The platform

A consequence layer that sits between intent and execution.

Causal Fields is model-agnostic and runtime-agnostic. One thin call wraps any tool your agent can invoke, so consequence-awareness becomes infrastructure — not a prompt.

Causal graph engine Expands each action into predicted effects and dependencies.
Risk scoring Severity, likelihood, and reversibility on every predicted branch.
Intervention policies Allow, gate to a human, or reroute — configurable per posture.
Full audit trail Every verdict is logged and replayable for compliance.

agent.ts

// Wrap any tool call with a Preflight check
const field = new CausalField({ posture: "balanced" })

const check = await field.preflight({
  action: "transfer 40,000 to vendor #A-19",
  context: agentState,
})

if (check.verdict === "BLOCKED") {
  return field.intervene(check)
}

await tool.execute() // only runs when safe

Why this becomes infrastructure

Every autonomous system will need a place to ask: “what happens if I do this?”

Compute has a layer. Storage has a layer. Identity has a layer. As agents gain the power to act, consequence becomes the next horizontal layer of the stack — and Causal Fields is building it.

Founding pilot

Put one real decision through Preflight.

A supervised 30-day engagement to configure, test and document one high-impact agent workflow.

Causal Fields Founding Pilot

490 € excl. VAT

One-time payment after the scope is approved. No subscription or automatic renewal.

Book the pilot Try the Lab first

Included in the pilot

1 operational use caseframed with your team and translated into decision constraints.
1 custom domain packcontexts, effects, thresholds and intervention rules.
Lab + API accessup to 500 supervised Preflight analyses during 30 days.
Launch and reviewa setup session, direct support and a final decision report.

Supervised pilot Causal Fields recommends and documents. It does not execute critical actions autonomously.

Pilot request

Describe the decision you need to secure.

We validate the scope with you before payment, then configure your first domain pack.

Your email app will open with the request pre-filled. No payment is taken now.