Follow the response from signal to effective intervention.

AGDA™ examines four connected capabilities: Detect, Escalate, Decide and Intervene. Together they form the SEDI intervention chain.

Scope first, then evidence.

We first define the system, the consequential pathway and the scenario. We then examine the evidence for recognising the need to act, reaching decision authority, making the decision and putting it into effect within the available time.

The assessment examines material dependencies and identifies the parts of the chain that constrain intervention. A policy fallback is not a measurement, and a walkthrough or a record of a previous drill is not a live operational test. Repeatable engine computation supports the assessment. Intervene remains responsible for evidence interpretation and the final conclusion. This page explains the basis and limits of the assessment; the underlying rubric and protected implementation remain Intervene IP.

The intervention chain. The chain is the control.

Detect, Escalate, Decide, Intervene. Four sequential stages running against the available time. The engine computes whether the cumulative chain completes before harm becomes irreversible. A full AGDA™ Assessment returns one of three states.

  1. Detect
  2. Escalate
  3. Decide
  4. Intervene
Each stage is necessary. None is sufficient. Intervention Readiness equals the weakest stage.
  • Before impact

    The chain completes in time under the observed conditions.

  • At risk

    The chain completes, but within a margin that operational noise can erode.

  • After impact

    The chain does not complete in time. Harm is irreversible before intervention can land.

Eight dimensions.

The four chain stages, plus the four conditions that determine whether each stage holds.

  • Detect

    Can the system surface the conditions that need attention, in time?

  • Escalate

    Does the signal reach an authority before the window to act has closed?

  • Decide

    Can the decision be made under pressure with incomplete information?

  • Intervene

    Can the action take effect before the consequence compounds?

  • Accountability

    Is responsibility real in practice, not only on paper?

  • Human oversight

    Is oversight informed and capable of altering outcomes?

  • Auditability

    Is the decision trail complete, reliable, and defensible later?

  • Capacity

    Are the people and authority present when the moment arrives?

Evidence quality constrains confidence.

Confidence cannot outrun the evidence. Policy, drill, production and stress evidence carry different weight before the engine propagates the chain.

Assumed confidence 1 / 5
Policy confidence 2 / 5
Drill confidence 3 / 5
Production confidence 4 / 5
Stress-tested confidence 5 / 5
Policy alone cannot carry a high-confidence intervention verdict. Green requires evidence the control has worked under real pressure.

A downstream stage cannot score more than one above the stage before it.

This follows from how the chain operates, not from a rule of thumb. A strong downstream claim on a weak upstream stage is not supported, because the downstream stage depends on what reaches it.

2.0 Detect
claimed 4.0 3.0 Escalate
claimed 4.5 3.5 Decide
claimed 4.8 4.0 Intervene
Detect measures 2.0 because signal is limited to nightly batch review. The chain ceiling propagates forward. Intervene is capped at 4.0, not the claimed 4.8.

Four positions the engine takes.

  • Coincidence is not a margin.

    An absence of incidents does not show that controls work. It may only show that they have not been tested.

  • Accountability cannot mask weak oversight.

    Naming an SMF4 who holds no halt authority assigns responsibility without capability. The engine assesses the pairing.

  • Evidence quality constrains confidence.

    Policy, drill, production and stress evidence carry different weight. High confidence needs evidence that has survived pressure.

  • Confidence must not exceed evidence.

    High confidence against weak evidence is demoted. Overconfidence is an assurance failure.

Signed SEDI result records, where supplied.

Signed SEDI result records are not a standard deliverable. Where one is supplied under an approved output-access arrangement, it can be checked against the signed commitments. The attestation envelope commits to the engine, methodology and rubric versions and to the input, output and manifest hashes. What the check does and does not establish, and what the signature does not cover, is on the verify page.

Example envelope · illustrative values 
attestationVersion 1.2
engineVersion      2.0.6
engineHash         17594a80…60355e50
transformVersion   N/A
catalogueVersion   1.0
methodologyVersion 1.3.0
rubricVersion      1.3.0
keyId              2026-05-02-d7e5468e
scoredAt           2026-05-14T09:47:12.334Z
inputHash          47a2e19c…103e4f8d
outputHash         91f4c02a…d4e7ad91
manifestHash       36b45cbe…3edc1c12
signature          M4Gb2ZqAfD…pLQX7w==

The values above are illustrative and show the shape of the envelope. A real signed record is produced by an assessment and commits to that assessment's engine version, input and output hashes. The published sample verdict is a constructed example and carries no signature.

How the verify path works

The documented measurement basis.

Four documents anchor the practice. None is published on this site; each is provided under NDA during diligence. There is no populated corpus and no outcome-validation evidence. AGDA™ is not validated against real-world outcomes.

  • AGDA™ Thesis

    The intellectual position. Eight falsifiable claims.

  • Failure Mode Catalogue

    Forty canonical intervention failure modes, referenced by identifier in a verdict.

  • Corpus Governance

    The governance mechanism by which validation cases would be authored, challenged and locked. It exists as a mechanism; no populated corpus exists yet.

  • Rater Protocol

    The protocol external raters follow when governed corpus evidence is supplied.

A sample verdict is published.

A constructed retail-credit example: reaching an authority holder may exceed the assumed six-hour window, and the required quorum takes at least seven days. It shows the shape of a finding, using no client data. It is not issued and not signed.

Discuss your system View sample verdict