fv-data-health

fv-data-health

Bad coverage means every downstream number is a lie. This skill's job is to say so, loudly, before a recommendation gets made on top of it — not to recommend anything itself.

Read shared/reading-fullvision-data.md before calling anything.

Steps

1. fullvision:query_view on view:health-identity-recon — what share of paying customers are linked to a tracked visitor. This caps attribution coverage: unlinked payers have no source, so every channel's revenue is understated.
2. fullvision:query_view on view:health-checkout-coverage — what share of checkouts carry a visitor cookie. Low coverage biases paid attribution specifically, because paid traffic converts faster and is more cookie-dependent.
3. fullvision:query_view on view:health-event-coverage — are product events arriving. A gap here means funnel and engagement views are incomplete for that window.
4. fullvision:query_view on view:ads-measurement-start — the first Stripe webhook charge. Spend before it is unmeasurable. Report the date; every ads skill needs it.
5. Grade and state the bias direction, not just the number.

Thresholds

Check	✅ ok	⚠️ degraded	🚩 broken
Identity recon (payers linked)	≥ 85%	60–85%	< 60%
Checkout coverage (cookie present)	≥ 80%	55–80%	< 55%
Event coverage (days with events / days in window)	≥ 95%	80–95%	< 80%

These are fixed. Do not adjust them at runtime to make a run pass.

What each grade means downstream

• ✅ — proceed, report the figure alongside recommendations.
• ⚠️ — proceed, but every ROAS is biased low by roughly the coverage gap. State the gap in the report and raise the minimum n by 50% for any write decision.
• 🚩 — stop. Do not recommend spend changes. The correct output is "fix tracking first", naming which check failed and what it breaks. A pause recommendation built on 40% identity coverage will pause profitable campaigns.

As a precondition

Every other skill calls this first and inlines the one-line result in its report header. A skill that reaches 🚩 aborts its own analysis and returns this skill's output instead.

Output

shared/report-format.md. Verdict is the worst of the three grades.

Refuse when

• Any health view is unavailable — an unknown coverage figure is not the same as a good one.