Open specification · JPS-1
The Judgment Preservation Standard™
Four measures of whether human judgment is still being exercised under AI assistance — defined the same way for a boardroom and a clinical setting.
Last updated:
Organisations adopting AI into consequential decisions are asked to govern something they cannot see. Model accuracy is measured constantly. Whether the humans in the loop are still exercising judgment is measured almost nowhere, and is usually inferred from the absence of bad outcomes — which is the one signal guaranteed to arrive too late.
This Standard defines four measures for that gap. It is published openly, for use and adaptation with attribution, because a measure that only one company can apply is not a standard.
Why one specification covers both a board and a clinic
The mechanism is the same in both rooms. A recommendation arrives, a human either engages with it or defers to it, and the record either shows which happened or it does not. Governance literature treats these as separate problems because the professions are separate, not because the mechanism differs. Defining the measures identically across both settings is what makes the comparison possible — and what allows a board and a clinical governance committee to speak the same language about the same risk.
The four measures
Override rate
How often does a human depart from the system's recommendation?
- Board setting. The share of AI-informed recommendations a committee alters or declines before adopting.
- Clinical setting. The share of decision-support recommendations a clinician modifies or rejects at the point of care.
Common misreading. A rate near zero is usually read as the system performing well. It is equally consistent with review having become ceremonial. The number alone cannot distinguish agreement from abdication, which is why it is never read on its own.
Concordance
When the human agrees, did they reason independently first?
- Board setting. Whether a director's own position is recorded before the AI-assisted analysis is circulated.
- Clinical setting. Whether the clinician's independent assessment is documented before the recommendation is surfaced.
Common misreading. High concordance is often reported as validation of the tool. Without an independent judgment captured first, it cannot be separated from anchoring, and the measure is uninformative.
Time-to-decision
Is the interval consistent with deliberation, or with acceptance?
- Board setting. Elapsed time between an AI-assisted analysis being circulated and a decision being taken.
- Clinical setting. Elapsed time between a recommendation being surfaced and an order being placed.
Common misreading. Falling time is normally reported as an efficiency gain. Falling time alongside falling override rate is the signature of review compressing, and should be escalated rather than celebrated.
Adverse outcome rate
Does the outcome record support what the other three suggest?
- Board setting. Decisions later reversed, written down, or found to have overlooked a material factor.
- Clinical setting. Patient safety incidents, adjusted for case mix, in AI-supported pathways.
Common misreading. This is a lagging measure. Waiting for it to move is the failure mode the other three exist to prevent. It confirms; it does not warn.
Provisions
- The four measures are read as a set. No measure in this Standard carries meaning alone. Each is defined so that it constrains the interpretation of the others. An implementation that reports one measure in isolation is not conformant with JPS-1.
- Independent judgment is recorded before the recommendation is seen. Concordance is only interpretable where the human position exists prior to exposure. Where the sequence cannot be evidenced, concordance must be reported as unavailable rather than estimated.
- Every consequential decision names an accountable person. The record shows who accepted, modified or rejected the recommendation, and on what basis. This produces decision provenance: a traceable chain from evidence to conclusion to accountable individual. A decision no individual owns cannot be reviewed and cannot be defended.
- Review happens on a cadence, not on an incident. The measures are examined at a defined interval agreed in advance. Reviewing only after an adverse event guarantees that the leading measures are read too late to act on.
- No threshold is asserted. This Standard deliberately publishes no target values. No threshold for any of these measures has been established, and a number invented for the appearance of rigour would be worse than none. Organisations set their own baselines and watch direction of travel.
- The Standard measures conditions, not quality. JPS-1 does not measure decision quality, and cannot. It measures whether the conditions under which good judgment is possible remain intact. That is a weaker claim, and it is the honest one.
Scope and limits
Stated plainly, because a standard that overreaches is worth less than none:
- It is not a clinical instrument. Nothing here is diagnostic, and nothing here is a substitute for clinical judgment, institutional policy, or regulatory requirement.
- It does not evaluate, certify or rank any AI system, vendor or product.
- It does not establish thresholds, benchmarks or pass/fail criteria.
- It confers no accreditation, and implies no endorsement by any regulator, standards body or professional association.
- It is offered for use and adaptation with attribution. It carries no warranty.
Status and attribution
JPS-1, released 2026-08-27. Authored by Dr. D. Ivan Young at Young Ethical Intelligence, Inc., and derived from Recursive Judgment Science™, the framework he developed. The evidence base is practice-based, qualitative and directional; no external validation is claimed, and none should be inferred. See the Research and Claims Statement.
Cite as: The Judgment Preservation Standard™ (JPS-1), Young Ethical Intelligence, Inc., 2026-08-27. Comments, adaptations and objections are welcome via the partnerships page; the measures are expected to change as they meet practice, and a superseding version will say so.
Background reading: protecting clinical judgment and decision authority for executives. Terminology is defined in the FAQ.