McPherson AI
Governance and independent verification

Your agent says the job finished. Observa checks what actually happened.

Observa is the accountability layer for action-taking AI systems. It discovers what your agents can do, maps those capabilities to governed meaning, captures execution evidence, and independently verifies supported outcomes against the systems that hold the real state.

The private beta runs in SHADOW mode. Observa observes, evaluates, maps, and verifies without taking control of production actions.

Request Private Beta Access Watch the Demo See public-safe proof → Docs and help →
Private beta open n8n + OpenClaw runtimes SHADOW_ONLY · No production control
See Observa in action

Watch an AI agent move from runtime activity to reviewable governance evidence.

Observa captures what happened, keeps runtime claims separate from verified outcomes, and shows what governance would have done while remaining in SHADOW. No production control at any point in this recording.

SHADOW ONLY · AUTHORITY NONE · ENFORCEMENT OFF

Product walkthrough

From a real agent task to a reviewable governance record.

An OpenClaw agent gets a real task and reaches outside the model. Observa discovers the real capability surface, a human approves the mapping and classifies it, and the boundary becomes a lifecycle record. Mapping is proposed and human-confirmed, never applied automatically.

Observa product demo · 43 seconds · has sound; nothing plays until you press play. An agent gets a real task and reaches outside the model, Observa discovers the real capability surface, and a human approves and classifies it. Discovery is automatic; meaning stays human-controlled.

Governance Analysis

What policy would have decided, while staying in SHADOW.

Available now in Hosted Observa for evaluated activity. A dedicated walkthrough of these states is in production; until then, here is exactly what the words mean.

  • WOULD_ALLOWAn enforcing policy would have permitted this action.
  • WOULD_DENYIt would have refused — and the action still ran, because nothing is enforced.
  • WOULD_REQUIRE_APPROVALIt would have held this for a human. Nothing was held.

Counterfactual by construction. Not every runtime action receives a decision; activity never becomes a decision on its own, and unevaluated activity stays visible as UNEVALUATED.

How to read these states →
Explore the beta Read the docs
The conceptual model

Identity, capability, authority, execution, reality.

01

Discover

Identify the agents, tools, and capabilities operating in the connected environment.

02

Map

Create reviewable mappings between capabilities and governed meaning. A mapping proposal is not approval, activation, or authority; a person reviews and classifies it.

03

Observe

Capture structured execution evidence without taking production control.

04

Govern

Evaluate activity against defined governance boundaries and keep the results reviewable.

05

Verify

For supported workflows and authoritative systems, independently determine whether the intended state actually exists.

The key distinction

Runtime completion is not the same thing as verified.

Agent or runtime

“Task completed.”

Observa

“Did the authoritative system actually reach the intended state?”

Independent verification applies where an appropriate verification contract or profile exists for the workflow and the authoritative system. Observa does not claim that every arbitrary workflow can automatically be verified.

Verification boundary

The result should never outrun the evidence.

If authoritative readback confirms the intended state, the outcome is verified. If it disagrees, the mismatch stays visible. If evidence is unavailable or uncertain, the outcome remains nonverified. Uncertainty is never silently turned into success.

What verification can do

  • Compare supported workflow outcomes with authoritative system state
  • Mark an outcome verified only when readback confirms the intended state
  • Surface mismatches between what was claimed and what exists
  • Keep unverified and uncertain outcomes clearly labeled

What Observa does not assume

  • That a successful runtime response proves the intended downstream state exists
  • That every tool, system, or workflow has authoritative readback
  • That missing evidence means no action occurred
  • That a readable report is a compliance or safety certification
Public artifacts

See the evidence structure in public-safe form.

These earlier artifacts show Observa’s reporting and evidence-presentation capability. Reporting is one Observa capability, not the whole product, and these artifacts do not validate runtime integrations or verification coverage.

Sanitized demonstration · PDF

Observa Audit Mode demo

A public-safe proof-of-work case showing how recorded activity can become a business-readable accountability report.

Open the demo PDF
Early public schema · v0.1

Observa Audit Mode schema

The case structure for status, confidence, evidence, flags, and the measured-versus-guidance rule.

View the schema
Observa private beta

Map one real workflow with us.

The Observa private beta is onboarding selected n8n and OpenClaw operators and builders. Founder-assisted, SHADOW mode, and honest about what it can and cannot verify.

Request Private Beta Access