PT Peter Tak

Research / Evaluation methods

Human-Centered AI Deployment Readiness Protocol

An interface can look ready before it is ready to build. Make the handoff decision with evidence.

A practical method for reviewing AI-generated prototypes: explicit requirements, observable states, recovery paths, and a separate record of the evaluator’s judgment.

v0.1-rc.2 Engineering handoff Yejun Tak · September 8, 2026
Review recordSynthetic example
Evaluator judgmentReady
Criterion statusNonready
1 unresolved critical issue

Hold for remediation.

A judgment is recorded. The evidence determines the next step.

01The approach

Separate the impression from the evidence.

What looks complete

Interface polish and functional completeness describe different properties. A convincing happy path can still omit a required state, a recovery route, or a critical requirement.

What can be demonstrated

Freeze the criterion. Collect the evaluator’s judgment before showing the reference key. Compare that judgment with documented evidence, preserving disagreements and missing information.

02Measurements

Five questions. Distinct answers.

Evaluate the reviewer and the artifact separately. There is no combined score, and high recall cannot override an unresolved critical issue.

01Reference-set defect recall
How many frozen reference defects did the evaluator correctly identify?
02False-ready acceptance
How often was a criterion-nonready artifact judged Ready? Report decision coverage, abstentions, missing judgments, and false holds on ready controls alongside this rate.
03Expected-recall gap
How far did expected recall differ from observed recall, in percentage points?
04Recovery coverage
How many applicable recovery scenarios were specified and walkthrough verified?
05Requirements-omission recognition
How many predefined omissions did the evaluator notice? Measure artifact requirements coverage separately.
03Assessment

From frozen brief to a recorded decision.

  1. Step 01

    Freeze

    Fix the artifact, task brief, reference set, and handoff criterion.

  2. Step 02

    Evaluate

    Lock findings, expected recall, and judgment before revealing the key.

  3. Step 03

    Reconcile

    Reveal the reference and adjudicate matches and disagreements.

  4. Step 04

    Record

    Calculate results, reconcile evidence, and record the owner’s decision.

Worked example / Synthetic

Walk through a service request.

A fictional specification with ten requirements, six recovery scenarios, and a separate reference key. Follow constructed findings through the completed workbook.

No real participant results or employer product data.

5 / 8Reference defects detected
3 / 6Recovery scenarios verified
Open the fictional artifactRead the explained calculations
05 / Design system

A shared language for the interface and its review.

This page uses the same typography, tokens, navigation, buttons, and cards as my portfolio. The reusable Peter Tak Design System also provides component specifications for constructing reference cases.

From component to evaluation evidence

The documented connection examines keyboard behavior, contrast, and confidence presentation. Component availability and a recorded correction are useful implementation evidence; they do not establish external adoption or protocol validation.

Read the connection and correctionOpen the reusable design system repository
06 / Research context

Method, version, and provenance.

The method relates explicit requirements and human judgments to evaluation evidence. Its NIST references provide context, not certification or endorsement.

NIST AI RMF 1.0TEVV-Athlon initial public draft

Cite this release

Tak, Y. (2026, September 8). Human-Centered AI Deployment Readiness Protocol: Engineering-Handoff Profile (Version 0.1-rc.2) [Evaluation protocol and materials]. GitHub. Versioned release.

Download BibTeX citation

Zenodo archival publication is pending. No DOI has been assigned.

Yejun Tak · CC BY 4.0 for original protocol documents and synthetic data; MIT for software. Development assisted by OpenAI Codex. Author review and external validation are separate steps.