Public Preview
Human-centered AI Readiness and Decision Protocol
Keep generation from outrunning understanding.
AI-assisted tools can compress the path from an intention to a convincing plan, design or implementation. H.A.R.D. starts with the result, reopens consequential choices, models the system truths and assumptions they depend on when needed, challenges conditions that could make them wrong, and connects the next human commitment to evidence.
Research motivation
Start with an important choice visible in the result. Decompress its purpose, alternatives, assumptions and tradeoffs. For consequential engineering choices, make the relevant source of truth, state, ownership, boundary, timing and failure model inspectable.
Then challenge the model: what would have to be true for this decision to be wrong? Separate a recorded earlier rationale from a new explanation, and keep unknowns unassessed until evidence exists. A well-supported existing choice can stay.
Decision compression is a motivating model for this preview, not an established causal effect. The proposed study has not established that H.A.R.D. improves decisions or defect detection.
In practice
Review the actual artifact revision. Record purpose, alternatives, rationale provenance, assumptions and tradeoffs without inventing creator intent.
When engineering deepening applies, externalize the relevant truth, ownership, state, boundary, contract, failure and time/ordering surfaces. Give consequential assumptions their own evidence and revisit lifecycle instead of burying them in rationale.
Name a bounded condition that could disconfirm the choice or assumption. If it is assessed, label the evidence as walkthrough, implemented or runtime-tested. Do not replace missing evidence with a familiar best practice.
Keep specified, walkthrough, implemented and runtime-tested evidence separate and bound checks to the exact decision context.
Accept, revise or keep the choice pending. Name the owner, smallest coherent next step and revisit trigger.
A separate structure view is optional review guidance, not an extra gate. Keep it separate from the proposed study design.
Choose the purpose of your review
A small team can inspect one artifact now. Add the six-gate review when deciding on an engineering investment, or an independent evaluation when measuring reviewer performance.
One person or a small team
Review choices, requirements and recovery in a plan, prototype or code artifact. Deepen consequential engineering choices only when their state, authority, timing or assumptions matter.
A plan can support a specification handoff. It cannot establish runtime readiness.
Start Artifact ReviewBounded engineering investment
Use QUICK-6 for low-risk work with a baseline, or Full Profile for moderate, high or unknown risk.
Stop at a failed or missing gate. The 15-minute target remains untested.
Open QUICK-6Open Full ProfileOptional evaluator research
Add frozen references, independent roles and locked judgments to examine reviewer performance.
Ineligible measures stay N/A with reasons. Agent results remain separate from human results.
Read evaluation rulesRecord the steps, time, failures and rework.
Tie the requirement to an actual need.
Define truth, authority, failure, timing and recovery where they matter.
Trace the requirement to the artifact and its check.
Count review, correction and escalation.
Apply the risk depth and name an owner.
Stop at a failed or missing gate. A good answer elsewhere cannot cancel it out. Without a measurable baseline, ROI stays indeterminate.
Look across the work
Use these four views to bring the right people into the review. They are discussion aids, not extra gates or separate scores.
Can people find their way?
Check labels, information hierarchy and competing choices against the task people need to complete.
What is true, and who may change it?
Follow source of truth, ownership, state transitions, boundaries, dependencies and the ordering or failure conditions that can contradict them.
What has been checked?
Distinguish working behavior from simulation. Link each important requirement to a check of the exact artifact.
Who handles the remaining work?
Name the people responsible for review, correction and support. Include their time when estimating savings.
A stage-bounded artifact result, or a separate engineering recommendation.
Choices, purpose, alternatives, assumptions, relevant decision surfaces, challenge evidence and remaining judgment.
A named owner, a resource limit and, for software work, the smallest coherent slice that can expose the important assumption.
The owner authorizes engineering separately; deployment requires separate evaluation.
The proposed study varies visual fidelity while holding content, behavior and defects constant. It does not establish that all AI-built work skips design steps. A separate structure view is a practice aid, not an unplanned intervention for that study.
Read the research question and limitsPractitioner feedback informed the protocol. Named credit appears only with permission.
| Contributor | Contribution |
|---|---|
| Hillel Glazer | Feedback on tracing artifacts to requirements and the tests or validation that demonstrate them, including reference material and form, fit and function. Attribution approved. |
To request attribution credit, contact takyejun@gmail.com or ytak47189@ucumberlands.edu.
Credit is not endorsement or validation. Other correspondents' identities and private comments are not published.
This is a Public Preview. Decision compression and implementation outrunning understanding are motivating models, not established causal effects. Practitioner correspondence informed refinement but is not controlled empirical validation. Software tests check implementation behavior. Usability, time savings, decision improvement and defect-prevention effectiveness remain unvalidated.
Interested in an external pilot or scheduling an official pilot walkthrough? Email takyejun@gmail.com or ytak47189@ucumberlands.edu with one workflow you would like to review.
Exact protocol 0.3-preview.1 · MCP 0.3.0rc1 · Skill/contract 0.3.0-rc.1. Existing commands, API names and HCAI criterion IDs are unchanged.
The earlier previews, rc.3 and all six rc.4 candidates remain unchanged. Past records keep their original names and versions.
Migration notes · Software test record · Download checksums · Preserved first preview
Tak, Y. (2026). Human-centered AI Readiness and Decision Protocol. H.A.R.D. Protocol 0.3 · Public Preview. Exact protocol 0.3-preview.1. No DOI is assigned to this preview. 10.5281/zenodo.22667623 identifies rc.3 only.
Original text and synthetic data: CC BY 4.0; software: MIT. AI-assisted development. Research Harness v3 is a separate project, not validation evidence. Source repository.