High-Risk AI Decision Readiness Score

01 Before a high-risk AI decision reaches a human reviewer, is it automatically triaged by risk level so the highest-risk decisions get the most scrutiny?

Yes -- every AI decision is scored and routed by risk level before review Some decision types are risk-ranked; others reach reviewers without triage We plan to implement risk-based routing but it is not live AI decisions reach human reviewers without automated risk ranking

02 Can you replay any past AI-assisted decision -- exactly as it unfolded, including the AI recommendation, the evidence the reviewer saw, and the final human decision -- without reconstructing it after the fact?

Yes -- every decision has a replayable audit envelope, accessible on demand Some decisions are logged but replay is incomplete or manual We are building replay capability but it is not yet in production We cannot replay past decisions in a structured way

03 When a human reviewer disagrees with the AI recommendation and overrides it, is the reason for that override captured in a structured, auditable format?

Yes -- every override is captured with a structured reason code and free-text field, linked to the specific AI recommendation Overrides are logged but without structured reason capture or not consistently We plan to add override reason capture Reviewer overrides are not systematically logged with reasons

04 Do you know, right now, what percentage of AI recommendations your human reviewers are accepting without modification -- and whether that rate has changed over the past six months?

Yes -- override rate is tracked per reviewer and role, with trend monitoring We have aggregate override rates but not per-reviewer or trend data We plan to track this We do not track reviewer override or acceptance rates

05 Do you have a way to measure whether your human reviewers are catching errors in AI output -- not just approving or rejecting, but actively detecting mistakes the AI made?

Yes -- we test error detection with known-error scenarios and track accuracy over time We review AI accuracy in aggregate, but we do not measure per-reviewer error detection We intend to build error detection measurement We assume reviewers catch errors; we do not measure it

06 Is there any way to distinguish a reviewer who thoroughly examined the evidence from one who rubber-stamped the AI recommendation in five seconds?

Yes -- we capture review time, evidence interaction, and modification depth per decision We capture time-to-decision but not evidence interaction or modification depth We plan to add review depth measurement All approved decisions look identical to us regardless of how much review occurred

07 After your AI system makes a visible mistake that a reviewer catches, do you measure whether that reviewer's behavior changes appropriately afterward?

Yes -- we track post-error behavioral adjustment and flag both under-correction and over-correction We review the incident but do not systematically track behavioral change afterward We plan to build post-incident behavioral tracking We handle incidents in isolation; we do not monitor downstream reviewer behavior

08 Could your reviewers make accurate decisions on high-risk cases if the AI system were unavailable -- and do you have data to support that claim?

Yes -- we have measured independent decision quality and it is within acceptable range We believe they could but we have not measured independent performance recently We plan to assess independent reviewer capability We do not know how our reviewers would perform without AI assistance

09 For AI decisions involving protected-class factors, adverse actions, or outcomes that could be legally challenged (denials, terminations, eligibility decisions), are domain-specific rules enforced automatically at the decision layer -- not just documented in policy?

Yes -- domain rules are wired into the decision routing layer; policy-relevant decisions cannot be auto-approved Some policy rules are enforced at the system level; others rely on reviewer judgment We plan to implement automated policy rule enforcement Policy rules exist in documentation but are not technically enforced at the decision layer

10 If a regulator asked you today for evidence that your AI-assisted decisions involve genuine human oversight -- not just a human in the loop -- how ready is that evidence?

Ready -- we have structured logs, override records, reviewer accuracy data, and audit-ready reports we can produce on demand We have logs but they would require manual compilation and interpretation to present as evidence We recognize the gap and are building toward audit-ready evidence We would need to reconstruct evidence after the fact; we do not have it ready

11 Do you monitor judgment quality trends across your reviewer population -- not just individual incidents -- to detect systemic drift before it becomes a compliance or liability event?

Yes -- we track reviewer population-level behavioral trends and have alerts for distributional shifts We monitor aggregate performance metrics but not behavioral drift at the reviewer level We plan to build organizational-level reviewer monitoring We respond to incidents; we do not proactively monitor for drift

12 Which best describes your organization's AI oversight goal for the next 12 months?

We need to get basic telemetry and JIS monitoring live before our next compliance review We need continuous automated calibration maintenance and an auditable IMP layer We are pursuing a third-party-verified oversight attestation for enterprise customers or regulators We are still assessing our oversight gaps and do not have a defined target yet

Nothing is sent anywhere until you choose to email it.

0 / 100

Nascent

DRS Readiness

JIS Readiness

Nascent Developing Operational Mature

Regulatory Deadlines

Today 2026-06-04

EU AI Act Annex III 2026-08-02

CO SB26-189 2027-01-01

Top gaps COHESION closes

COHESION Compliance Evidence Report

Organization: [Your Organization] Period: [Q2 2026]

Field	Value	Status
oversight_score		computed
drs_events_total		computed
jis_composite	--	requires live stream
band		computed
top_gaps		redacted

This is what your compliance officer receives. Every AI-assisted decision. Every reviewer. Audit-ready.

Get the sample evidence report (PDF)

DRS vs. JIS: DRS scores the AI decision's risk level and routes it to the right reviewer. JIS scores the human reviewer's judgment quality. Both are required for provable oversight. This diagnostic estimates your readiness on both axes.

Send my score

See what COHESION instruments Talk to the founder Get the sample evidence report (PDF)

COHESION saves humanity by keeping human judgment alive in the age of AI.