Skip to main content
ControlFrame

Agent orchestration · governed evidence operations

The audit room gets an agent team.Not an ungoverned chatbot.

This page proves two boundaries with an event-fed synthetic replay: six specialist roles execute one governed evidence record inside declared input, output, and tool limits — and every decision that creates accountability stays with an authorized person.

Specialist roles
6
Reference artifacts
4
Automated releases
0
Public mode
Event-fed

01 · Stream the proof

Watch each event arrive, then inspect what its agent cannot decide.

Synthetic reference data · eight digest-linked envelopes · production collection requires a configured private runner
synthetic reference replayGoverned six-role contract replay
GET proof readyRUN-8837Download trace

Release authorityAuthorized human held
One synthetic record moves through six bounded roles; every output stays attached to its source, digest, review state, and release policy.
04advisory

Control mapper

Preserves native authority and proposes source-cited reuse candidates.

Input
CF-UI-03 + accepted evidence
Output
2 native mappings + 4 review candidates
Authority boundary
A candidate never changes a target control status automatically.
role event contractMAPSET-EV2048
  1. 01native.mapping.preserved
  2. 02candidate.support.proposed
  3. 03review.requirement.attached
TransportGET proof ready
Envelope
0/8
Decision
pending
sha256:awaiting-event
Artifacts
4
Mappings
6
Auto-releases
0
Synthetic reference · Private runner not activated · production collection requires an approved tenant configuration.0 automated releases

02 · The execution contract

Parallel intelligence. One controlled sequence of record.

  1. 01Scope

    Declare the evidence contract

  2. 02Collect

    Execute the approved plan

  3. 03Validate

    Check shape, digest, and custody

  4. 04Map

    Propose source-cited reuse

  5. 05Challenge

    Test sufficiency and sensitivity

  6. 06Package

    Recompute gates and hold release

Specialist reasoning can fan out. The evidence record does not: every artifact, validation, mapping candidate, finding, and decision returns to the same governed chain.

03 · Where autonomy ends

Automation stops before accountability begins.

Human release authority is a product control, not a disclaimer.
  1. 01
    Accept or reject evidence

    Agents can prepare a recommendation; a scoped reviewer records the decision.

  2. 02
    Approve redaction and exceptions

    Sensitive-data treatment and material exceptions stay attached to accountable people.

  3. 03
    Authorize package release

    Eligibility can be computed automatically. Export authority is never delegated to the agent team.

04 · Runtime truth

The interface tells you what is shipped, replayed, and still gated.

Shipped primitivesRepository implementation

Typed orchestration, tenant-scoped evidence intake, digest and custody records, review separation, fail-closed package gates, and offline receipt verification.

This public surfaceEvent-fed synthetic proof

A deterministic GET-only event stream drives eight versioned envelopes across six bounded roles and the held terminal state for RUN-8837. It proves the public contract; it is not a claim that a customer production runner is active.

Production activationPrivate runner not activated

Provision the private runner, enroll approved targets and identities, configure independent secrets and storage, then pass authenticated tenant-scoped smoke evidence.

05 · The difficult layer

The moat is the proof protocol—not the chat interface.

01

Plan before execution

Source, expected proof, target, freshness, tool boundary, and review policy are declared before work begins.

02

One governed record

Artifact bytes, digest, provenance, sensitivity, mappings, findings, and decisions stay connected.

03

Separation of authority

Agents execute and advise. Authorized people own sufficiency, exception, acceptance, signature, and release.

04

Proof that travels

Eligible records carry manifests and verification material beyond the ControlFrame session and portal.

Take the next evidence-backed step

A credible AI system should show where its authority stops.

Governed Agent Orchestration | ControlFrame