Evidence Development Program

What the research shows, and what remains unproven.

The research examines reviewer identification of Decision Reconstruction Risk in constructed records. Its findings do not establish effectiveness in operational records.

Read this first. The principal detection results come from constructed, AI-generated records with known answers. They do not establish effectiveness on real workplace, investigation or administrative records. The human-reviewer reliability criterion and the cross-vendor reproducibility criterion have not been established.

Explore the evidence by purpose

Select the level of detail needed for your review. The summary is intentionally short; original methods, records and limitations remain available.

01 · Findings and limitations

What was measured?

Expert detection accuracy on constructed records, differences between reviewers, agreement measures and the limits on interpretation.

Read the research findings and limitations
02 · Study methods

How were the findings produced?

Study designs, participant and record sets, scoring rules, pre-registered criteria, measurement choices and analytic constraints.

Examine the methods and criteria
03 · Evidence and provenance

Can the underlying claims be traced?

Research chronology, documentation and available aggregate datasets. Restricted participant and source materials are not published here.

Open the public evidence chronology
04 · Open questions

What must still be tested?

Field performance, individual reviewer dependability, construct validity and consistent application in real organizational workflows.

Review the open research questions

Evidence at a glance

The detection study involved 16 independent international experts from 11 countries across five continents, evaluating constructed AI-generated decision records against a held-out reference key. Their expert participation supports a defined research finding, not certification of JRS or proof of effectiveness in live organizational records. Read the study and its limitations.

83.9%Mean expert detection accuracy against a verified key on constructed records. Detection subset: 16 independent international experts, 384 graded reads; 95% participant-level CI 72.7% to 95.1%.
36 reviewersCompleted a full 24-record set across 16 countries. This broader completion group must not be confused with the 16-expert detection subset.
58 reviewersGraded at least one record across three studies. A participation count is not an accuracy or validation estimate.

The detection threshold was met on its pre-specified analysis. A separate reviewer-agreement criterion was not met because the confidence-interval lower bounds did not clear the required threshold. Cross-vendor raw agreement does not establish reproducibility. See the study-specific documents for estimators, denominators and limitations.

Separate reviewer-agreement sample

The reliability analysis used 104 determinations from 22 analyzed raters, with 10 multi-rater records. Gwet’s AC1 was 0.739 in the invited group and 0.623 in open enrolment. Both point estimates exceeded 0.61, but neither required confidence-interval lower bound exceeded 0.41. The pre-registered two-part criterion was not met.

These are not additional observations from the 16-expert detection panel. They do not establish individual reviewer dependability or operational validity. See analysis and uncertainty and study design.

International research and practitioner contributions

Different Minds. Fewer Blind Spots. Stronger Evidence.

JRS has been examined through research involving international practitioners, methodological scrutiny, and employment and public-records pilot work.

That work has produced findings, questions, and limitations worth examining.

It also points to a concern shared across professional settings: a consequential decision should not become impossible to understand simply because the original decision-maker is no longer available to explain it.

The research contributions below have different roles, samples and evidentiary limits. International participation and publication do not themselves establish operational effectiveness, institutional adoption or certification.

Research participation

Documented: 16 independent experts from 11 countries contributed 384 graded reads in the defined constructed-record detection study. The separate reliability sample and wider participation totals are not interchangeable.

Read sample sizes, findings and negative results.

Published professional discussion

Documented: A coauthored compliance article examines AI-drafted employment records and evidentiary support. Publication is a forum for scrutiny, not validation or endorsement by the publisher. Any future scholarly article must be linked with its exact publication status and study limitations.

Inspect the published article record.

Verified use and endorsement

Not established here: Research participation, publication, shared links and anonymous support clicks do not prove routine practitioner use or institutional approval. Named adoption and attributable endorsements require direct evidence, proper scope and permission.

Examine unresolved field-validation questions.

01 · Methodological scrutiny

Questions about the measurement method

Ubayet Hossain contributed documented review of the reliability methodology, including Gwet's AC1 and the proposed decision criteria. The separate reviewer-agreement study did not meet its pre-registered confidence-bound criterion. Methodological contribution is distinct from evidence of validated field performance.

Examine study design and reliability criteria
02 · International practitioner examination

Different readers examining defined records

Practitioners from multiple countries examined constructed records under specified study conditions. The expert detection subset, full-set completers and wider participation total have different denominators and findings. International participation broadens the questions asked; it is not a representative global validation sample.

Review findings, cohorts and limitations
03 · Domain-focused pilot contributions

Employment and public-records questions

Tanvi Pokhriyal contributed first-author work to the employment-records study, and Stacyann Young contributed first-author work to the public-records study. These exploratory contributions connect documentation review with distinct professional contexts. They do not establish routine workplace performance or independently verified operational outcomes.

Inspect the public research evidence chronology
04 · Published professional analysis

Bringing the documentation issue into professional discussion

Hekim Colpan and Phillip Wikes coauthored a published compliance article examining AI-drafted employment records, repeated subjective language and the need to trace judgments to contemporaneous evidence. Published commentary extends professional discussion; it is not an additional validation study or publisher endorsement of JRS.

Read the published coauthored analysis below

Named contributions are attributed to their documented research or publication roles. They must not be interpreted as organizational endorsements, unrestricted promotional permissions or evidence that all participants support either public initiative. Individual affiliations and restricted pilot records are not disclosed here.

Verifying resource distribution and use: A future example of an organization sharing or applying a JRS resource should be identified only when the exact resource, date, independently checkable evidence, scope of use and permission to name the organization are documented. A link placement is not proof of practitioner application, institutional endorsement, or improved records. No unverified adoption or endorsement is claimed here.

From research questions to two public initiatives

The international statement above is an invitation to examine a record from different perspectives, not a measured outcome. The free JRS resource library provides practical references for reviewing evidence, chronology and reasoning. The Right to Know Why addresses people affected by decisions. Decisions You Can Defend serves professional reviewers. Neither initiative creates missing evidence or proves that a decision is correct.

Affected-person perspective

The Right to Know Why

Read the Decision Reconstruction Risk (DRR) explanation and free educational references for people affected by decisions. The initiative expresses a public-interest principle, not a new legal or disclosure right.

Open free public-interest resources
Professional review perspective

Decisions You Can Defend

Access free investigator guides and the seven-point record check for examining missing support, chronology gaps and unsupported conclusions. A reviewable record does not establish that the decision is correct, lawful or legally defensible.

Open free professional resources

These two educational pathways are distinct from the empirical research. An on-page comprehension self-check is not a scored comprehension study; a shared resource link is not verified practitioner use; and voluntary support clicks are not proof of institutional adoption or improved record quality. How optional initiative support is counted.

Supporting research records

The full technical detail is preserved for reviewers who need to examine the complete record.

The tagged reading copies have document structure and headings but have not been independently certified for PDF/UA or screen-reader use. The original PDF editions remain unchanged. The public study pages do not confer certification or regulatory compliance, provide access to confidential datasets, or establish effectiveness in real-world operational review.

Professional commentary and applied examples

The following external articles discuss documentation integrity in HR and compliance work. They are authored commentary, distinct from the empirical research and its limitations described above.

Corporate Compliance Insights · September 29, 2026

That AI-Drafted Termination Memo Could Become Evidence

Coauthored by Hekim Colpan and Phillip Wikes. Examines how AI-drafted employment records can obscure source evidence, repeat subjective characterizations across files, and complicate independent examination of the reasons for a decision. Published professional analysis, not an additional JRS validation study.

Read at publisher ↗
SCCE · CEP Magazine · November 1, 2026

When the record cannot speak for itself

By Phillip Wikes. Examines source traceability, the limitations of template compliance, and documented independent review.

Read at publisher ↗

Appearance in a professional publication does not establish JRS operational effectiveness or confer independent approval.

What field application would need to establish

Constructed-record detection, independent reviewer agreement, software engineering checks and actual workplace use require separate evidence. Review the current evidence boundaries and an inspectable field-observation procedure. No real-world operational effectiveness claim is made by these constructed-study results.

Publication and research citation controls

For corrections or a methodological inquiry, use the public Contact form. Free practitioner resources remain available without submitting contact details.