The public artifact is not presented as if it arrived fully formed. This record shows the product mistakes, design changes, and engineering decisions that shaped it.
The first version asked users to describe reviewer roles, evidence, authority and escalation in professional language.
What was wrongYou effectively needed to understand AI governance before you could use an AI-governance tool.
The tool became an eight-dimension quick check with a deeper protocol mode, prioritized weaknesses, and browser-local processing.
What improvedPeople could get useful first value without writing an essay or creating an account.
The copy was rewritten for a range of education and domain backgrounds. Expert concepts stayed in the system; the first interaction stopped requiring expert vocabulary.
ExampleReviewer competence
Does the reviewer know enough to spot a bad answer?
001 moved from a generic portfolio shell into an exhibition worktable. 001-B became a screening-room object rather than an embedded animation demo.
WhyThe containing portfolio should explain the work without making the usable product carry the burden of explaining how it was made.
Evidence provenance, source freshness, accessibility acceptance, test fixtures, B-Web/B-LinkedIn parity, lineage proof, privacy/security, performance, and technological-ceiling decisions became explicit release gates.
Why“Looks finished” is not the same as being inspectably finished.
The hidden architecture can remain rigorous. The visitor should encounter only the complexity they need at that moment.
The distance from v0 to v4 was not a sequence of refinements to a working idea. v0 failed for a structural reason rather than a cosmetic one: it asked the user to supply the very analysis the tool existed to produce. No amount of copy editing would have rescued it, and recognising that early is the difference between a rewrite and a redesign. Every later version is a consequence of accepting that the first framing was wrong rather than merely rough.
The second lesson is that the audience question kept moving after it appeared to be settled. v1 solved the expert-vocabulary problem for people who already worked adjacent to governance, and it was only at v2 that it became clear the tool was still assuming a professional context the reader might not have. Audience assumptions tend to surface one layer at a time, and each layer is invisible until the one above it is removed.
The third is that a build can be finished as a product and unfinished as a public object. v3 and v4 changed almost nothing about what the tool does; they changed what surrounds it and what has to be true before it can be published. That work does not improve the artifact for the person using it, which is precisely why it is the part most often skipped, and why it is recorded here instead.