Building a decision system people can audit.

The call

Build the evidence and release system before selling the marketplace.

A credential page can look complete while quietly mixing exam eligibility with recommended experience, a national occupation wage with credential-holder pay, or a dated provider price with live availability. RoleMath’s central product decision was to separate those claims before adding commercial behavior.

Decide what training is justified first. Keep provider sourcing, consent, and commercial activation behind a separate boundary.

The inherited problem

Certification decisions change over time. Exam versions retire, renewal routes change, member and nonmember fees diverge, and a vendor’s recommended background is not necessarily a registration gate. At the same time, acquisition pages can overstate salary, hiring, provider coverage, or “worth it” conclusions when the evidence supports only occupation context.

The system therefore needed two things at once: a decisive reader experience and a harder evidence boundary underneath it.

What I built

1. A temporal credential model. Current exam identity, checked dates, lifecycle transitions, fees, eligibility, renewal, preparation, conflicts, and known unknowns stay separate.
2. A decision contract. Strong pages begin with a recommendation, fit and non-fit boundary, one alternative, and one useful next action.
3. Claim-class safeguards. Occupation statistics describe occupations, not credential holders. Employer samples require a complete tuple or stay suppressed. Unverified relationships do not render publicly.
4. A gate-delegated release system. Deterministic checks and fresh adversarial review protect a wave without manufacturing human approval or forcing circular per-page signoffs.
5. An open-data boundary. Public facts, sources, dates, and limitations are downloadable; private scoring, review state, and commercial matching remain excluded.

A failed release can be a successful control

A controlled attempt to regenerate the production substrate accounted for all 310 protected public routes as 161 ready and 149 hash-void. Because 149 routes were not release-ready, the release was aborted. Production stayed on the known-good deployment. That result was not page-production progress, but it proved the release control would fail closed instead of silently demoting or changing public content.

First acquisition evidence—without inflating it

The first complete Search Console baseline covered June 19 through August 8, 2026: 3 clicks and 4,034 property impressions. Search Console exposed 520 query rows representing 2,391 impressions; anonymized queries account for the gap. Raw government-source filenames generated 1,240 exposed-query impressions, certification-decision queries 706, salary intent 107, and provider intent 81 with one click.

Page impressions are not additive to the property total when more than one RoleMath URL appears in the same result. The useful conclusion is not “traffic achieved.” It is that raw-source and salary visibility must be separated from the certification and provider decisions RoleMath intends to serve.

What the work proves

• Product judgment across research, data contracts, editorial UX, acquisition measurement, and release operations.
• Ability to convert messy vendor and public data into auditable decision infrastructure.
• Commercial discipline: training-price observations are useful without implying provider consent, inventory, willingness, or a relationship.
• Operational restraint: the live site stayed on a verified 310-route floor while experimental products moved to a separate public sidecar.

What it does not prove

This checkpoint does not prove unique visitors, qualified demand, conversion, revenue, provider relationships, employment outcomes, or certification-caused pay. The Fit Check is a bounded decision aid, not a readiness, salary, hiring, or pass-probability model.