Emergences AI, Inc.

Assessment Model Card

Version: 2026-07-03
Effective: July 3, 2026
System: Candidate assessment
Risk tier: High-risk (employment)

This Model Card describes the AI system NeoWork uses to evaluate candidates, so employers, candidates, and reviewers can understand what it does and does not do. It is written to be read by non-specialists as well as technical and compliance reviewers.

What the system does

The system evaluates a candidate's work on a realistic task and produces a descriptive report of how effectively they worked — with particular attention to how well they work with AI. It is a decision-support tool: it does not decide who is hired and does not output a “hire/no-hire” label.

Intended use and users

2.1Intended users. Employers and recruiters evaluating candidates for roles, who apply human review and make the hiring decision.

2.2Intended use. As one input into a human-made employment decision, alongside other evidence.

2.3Out-of-scope uses. Automated hire/reject decisions; using a score as the sole basis for a decision; ranking or filtering by protected characteristics; any use prohibited by the Acceptable Use Policy.

Inputs

The system reads observable signals from the assessment session — the candidate's screen recording, the on-screen content, the software and tools used, input behavior, the submitted files (the deliverable), and the process/event timeline. It does not take, request, or infer protected characteristics.

Outputs

Output	What it means	Scale
AI Quotient	Result level — what the candidate shipped with AI, graded against the same task done by AI alone. Reported by domain.	0–100
Competency	Process level — the AI-independent judgment (how they reason, verify, frame, and self-regulate) that holds as tools change.	0–100
Construct scores	14 constructs across 3 competencies, each supported by cited evidence from the work.	0–5 (or “insufficient evidence”)

Every construct score is tied to specific evidence. Where the evidence is insufficient, the system records “insufficient evidence” rather than a low score.

How it is scored

Two-stage scoring: first identify observable evidence, then combine that evidence to infer a level — no “overall impression” judgments.
Explicit criteria: every rubric reduces to a defined observable variable with inclusion/exclusion thresholds.
Three-valued logic: each criterion is confirmed-true, confirmed-false, or undeterminable — the last is distinct from a failure.
Unobserved is never penalized: a behavior that was not observed is flagged for human review, not scored as absent.
Multi-source corroboration: no single signal is decisive; the submitted deliverable wins on conflicts about the result.

Human oversight

The employer reviews the report and makes the decision. Evaluation logic is authored and changed under review. Low-evidence and conflicting signals are surfaced for a human rather than resolved silently.

Limitations

Results describe performance on one task and may not generalize to every aspect of a role.
Assessment of work done with AI depends on the quality and completeness of the captured session signals.
Like any AI system, the underlying models can make mistakes; this is why evidence is cited and human review is required.
The system is not a measure of a person's worth, intelligence, or fixed ability.

Fairness and evaluation

The design choices above — evidence-based scoring, not penalizing the unobserved, excluding protected characteristics and their proxies — are fairness controls. We assess the system for disparate impact and publish a summary when it is independently audited. [Measured performance and demographic-parity results to be added from the current validation and audit.]

Data, privacy, and versioning

Inputs and outputs are handled under the Candidate Privacy Notice and Privacy Policy; sub-processors (including AI model providers) are disclosed. This card is versioned; material changes to the assessment update the version and are noted in the Legal Updates archive.