How our valuation engine evolved
Four generations of pipeline, one goal: a declared value an auditor can defend. Pick a version above, then scroll — the story unfolds as you go.
MAVEN PROOF FILE E4-2214
MAVEN Proof 4.0 doesn't estimate — it builds a case. Follow a single donated item from the moment it's scanned to the moment a defensible value is declared, and watch what it takes for the engine to say it's confident.
From the Yiddish meyvn, rooted in the Hebrew mevin — “one who understands.”
A maven is not merely knowledgeable — a maven is the trusted expert others turn to when the answer has to be right. Our engine carries the name because it behaves like one: deep expertise in medical products, applied with judgment.
4.0 adds the second word. A maven's opinion is only as good as their evidence — so MAVEN Proof is required to show its work: every conclusion arrives with the sources, math, and rejections that prove it.
Intake — the case opens
Sealed surgical stapler reload box — three photos, one barcode.
A run record is created the instant the scan lands. From this point forward, everything the engine does — every query, fetch, acceptance, and rejection — is written to the case file.
PHASE 1
Before a single search runs, the engine establishes exactly what it's looking at. Photos, OCR text, and the barcode are fused into a Canonical Identity Record — and every field carries provenance: what proved it, and how strongly.
PHASE 2
The engine searches the way an investigator would: the GTIN first, then REF + manufacturer, then descriptive queries — across multiple engines. Every query and every attempt is logged. If a provider throttles us, that's an infrastructure event, never evidence about the product.
PHASE 3
38 candidate sources surface. Each is matched against the canonical identity — identifiers first, visuals second. A conflicting REF or UDI is a hard rejection: no amount of visual similarity can promote a wrong part number.
EXACT
LIKELY
PARTIAL
REJECTED
PHASE 4
Surviving candidates get their pages actually read — fetched under a global concurrency governor that respects provider limits. Each page is proven field-by-field against the canonical record: manufacturer ✓, REF ✓, GTIN ✓, packaging ✓. A page we were never allowed to read proves nothing, in either direction.
PHASE 5–6
Every validated source is classified on two separate axes: authority over what the product IS, and authority over what it SELLS FOR. A manufacturer page with no price is still Tier-A identity evidence. Two storefronts owned by the same company count as ONE independent observation. Prices are extracted only when visibly on the page — and only with an explicit per-each or per-pack basis.
PRICED SOURCES
INDEPENDENT GROUPS
PHASE 7
The priced pool converges: listings sold as a box of 3 are normalized to per-each with documented math — $937.50 ÷ 3 = $312.50. Price dispersion is 1.4×, well under the 3× adjudication threshold. The target quantity is evidenced (box of 3, explicit marking), so the line total is allowed to resolve. Confidence is then computed under hard structural caps — 95+ requires at least three independent pricing sources. This case has five.
PHASE 8
The case closes with a full traceability bundle: every accepted source with its URL and proof, every rejected source with its reason, every query, every fetch attempt, every normalization step, and the exact rule that produced each number. An external auditor can replay the entire reasoning chain.
Per each
Evidenced quantity
Declared line value
Forensic confidence
5 independent pricing groups · Tier-A identity proof · evidenced quantity basis