npbuilds Product Verifier — The Inspector
Action skill: per-finalist hard checks. AI shopping assistants demonstrably skip these — 18-34% of their recommendations point to discontinued "zombie" products, specs get hallucinated, and scam merchants poison results with cloned listings. No finalist reaches emptor's evaluation matrix unverified.
Description
Four checks per finalist: claim verification, availability, safety/reliability history, merchant trust. Claim verification reuses source-triangulator (research domain) at quick depth — standard depth for claims that decide must-have compliance. Source weighting follows source-trust-atlas; paywalled Tier 1 evidence on critical claims routes through paywall-strategist before being declared unverifiable.
Input
| Parameter | Required | Notes |
|---|---|---|
finalist |
yes | Product name + exact model/SKU if known |
claims_to_verify |
yes | The spec/marketing claims that matter to the requirements spec (esp. must-have compliance) |
depth |
no | quick / standard (default) / deep |
Process
- Pin the exact SKU. Model-year revisions and regional variants invalidate reviews silently; record which variant every piece of evidence applies to.
- Verify claims. For each claim: route to
source-triangulatorwith the claim, type, and the atlas-derived source chain. Manufacturer copy is the claim, never the evidence. Must-have-deciding claims get standard depth and a confidence tag each. - Check availability. Confirm the product is in production and purchasable today: manufacturer product page live, ≥2 reputable merchants stocking it, no "discontinued/replaced by" notices. Record status —
in production/discontinued (successor: X)/limited stock/unclear— with a check date. - Check safety & reliability history. Recalls (CPSC/NHTSA/FDA as applicable), known failure modes from Tier 1 reliability data, firmware/support status for connected products, warranty terms.
- Trust-check purchase merchants. For each where-to-buy candidate: domain age and reputation, marketplace seller vs first-party, return policy, price plausibility (a 40%-below-market price on a fresh domain is a scam signal), FTC/consumer-protection history. Verdict per merchant:
verified/caution/avoid. - Report with one block per finalist; every fact date-stamped and cited.
Output
PRODUCT VERIFICATION — <finalist> (<exact SKU>)
Claims:
- "<claim>": Confirmed|Likely|Speculative|Contested|Unverifiable because <evidence + [N]>
Availability: <status> (checked <date>) — <evidence>
Safety/reliability: <recalls, failure modes, support status, or "none found — searched X, Y">
Merchants:
- <merchant>: verified|caution|avoid — <reason>
Must-have compliance: pass | fail (<which must-have>) | unverifiable (<which>)
A finalist whose must-have compliance is fail is eliminated; unverifiable caps the brief's TOP PICK confidence at Speculative if that finalist wins.
Error Handling
| Failure | Response |
|---|---|
| Spec claim only on manufacturer site | Tag Unverifiable-pending; try independent teardowns/tests before giving up |
| Availability unclear (regional, fluctuating) | Report unclear with what was checked; never assume in-stock |
| All merchants are marketplace third parties | Verdict ceiling caution; say why |
| Paywalled Tier 1 review decides a must-have | Route to paywall-strategist first |
Scope Boundaries
Handles: facts about a specific product and its merchants.
Does not: score review authenticity (review-forensics), rank finalists (evaluation phase), or decide requirements (needs-elicitor).
Related Skills
- Depends on
source-triangulatorandsource-trust-atlas; verdicts feed emptor's VERIFICATION & AUTHENTICITY and WHERE TO BUY sections.
Learn Block
Next, learn recommendation-brief — how verified evidence becomes an auditable recommendation.