JA-ES v0.6 Interactive Reference Demo

Waiting for evaluation

May this case proceed automatically?

The left side is locked governance. The right side contains editable demo inputs that simulate source-system facts.

Passing every condition proves this case satisfied the authority that was claimed for it. It does not prove the boundary itself was appropriate, or that this population should have been eligible for automation at all. Whether the Envelope classified this case correctly is a separate human judgment under the Assessment and Evidence Guide.

Approved boundary — read only

Demo applicant facts

Score Requested amount AI recommended action Score age (hours)

Fraud flag present Thin-file indicator present Recent adverse action present Manual referral present

No case has been routed

What did the human do?

No human review is currently required. Evaluate a case that fails at least one boundary condition.

No sealed evidence yet

What evidence now exists?

Choose a sealed record

Can the evidence still be trusted?

Choose a generated record, alter one protected value, and compare it with the sealed baseline. This edits a test copy, not the original record.

Record to verify Protected value

PASS

Sealed hash

Recomputed hash

What these checks establish, and what they do not. A PASS means the record is well-formed, unaltered, internally consistent, and correctly chained. It does not establish that the reasoning was sufficient, that the boundary was appropriate, or that the organization is JA-conformant. A record can pass every check above and still fail human sufficiency review. Schema-valid is a machine result; JA-conformance is a human determination under the Minimum Control Standard and the Assessment and Evidence Guide.

Illustrative Guard diagnostic

What becomes visible across decisions?

A non-functional population view showing how structured Judgment Records could support oversight at scale.

SIMULATED DATA

Demonstration only. The figures below are hypothetical outputs across 12,486 decisions. They are not calculated from the single walkthrough record and do not represent a validated scoring model or a finding of substantive decision correctness.

Judgment Records analyzed

12,486

AI output accepted

78%

AI output modified

14%

AI output rejected

8%

Tier B reasoning used

31%

Complete ownership

100%

Tier A reasoning distribution

Consistent with policy

54%

Context not visible

17%

Data-quality concern

11%

Contradictory evidence

8%

Legal or regulatory constraint

6%

Diagnostic signals requiring assessment

Potential automation biasThree reviewers accepted more than 99% of outputs; their median review time was 7 seconds and 82% of records used the same Tier A reason.

Workflow or model signalOverrides citing “relevant context not visible to the system” increased 44% after model version 3.2.

Reasoning sufficiency concernSix percent of high-risk decisions used generic narratives that did not connect the selected reasons to the final outcome.

Positive control signalSeeded cases designed to require rejection were rejected 94% of the time.

Completeness depends on reconciliation. A failed evaluation creates a work item, but no Judgment Record exists until the human submits one, so a missing record is itself the finding. Guard reconciles Judgment Records against source-system routing events to identify referred cases that never produced a completed human decision.

Guard identifies patterns for human assessment. It does not mechanically determine whether an individual decision was correct or whether an organization is JA-conformant.

Who is authorized to let the machine act?