Skip to content
Methodology

Evidence has to earn the conclusion.

Commerce data is full of statements that look like facts: a specification nobody signed, a stock figure nobody promised, a certification nobody produced. Software that treats those as inputs produces confident output and transfers the risk to you.

Our method is the opposite discipline. A claim is admitted only when a source in a position to know has stated it about the exact subject, recently enough to matter. Everything else is recorded as unresolved or blocked — and stays that way in the output.

Nine principles

How a claim becomes admissible

01

Source binding

Every admitted claim carries the source that stated it. A conclusion with no traceable source is not a weaker conclusion — it is not a conclusion at all, and it does not enter the brief.

02

Exact subject identity

Evidence about a similar product is not evidence about your product. Before anything else is assessed, the review establishes whether each source describes this exact item and variant — a failure that can invalidate product research.

03

Source authority

Authority is judged per claim, not per document. A supplier may state its own price; it cannot establish a marketplace's requirements. The same page can be strong evidence in one line and worthless in the next.

04

Freshness

Evidence has an as-of boundary. Current exact-unit availability can be assessed when qualified, sufficiently current evidence supports the exact product and context. Prices, availability and routes can change; the review provides no future-stock guarantee or continuous inventory monitoring.

05

Missing proof is output

Every material question is accounted for as supported, conditional, unresolved with a reason, or outside the agreed scope. Missing proof is distinguished from a question the review cannot answer, so next steps address the actual gap.

06

Conflicts are preserved

When sources disagree, the conflict and permitted evidence references remain visible. We do not silently choose a winner; raw source statements are not automatically disclosed.

07

Decision traceability

The brief is readable backwards. From the outcome you can reach the claims, from the claims the sources, from the sources their age and authority. You should be able to defend the decision to someone who was not in the room.

08

Bounded economics

Economics considers supported revenue and direct-cost components for the agreed scope, including qualified freight and fee evidence where available. Known-cost contribution is shown with its coverage and missing inputs; it is not universal net profit, ROI or guaranteed margin. Customer-entered amounts remain unverified until supported by evidence.

09

Deterministic where possible

Matching, extraction, freshness, and consistency checks are rules. Rules are auditable and repeatable, so they carry the parts of the work that must not vary between two reviews of the same subject.

The binding test

Four questions per claim

Every candidate claim passes through the same four gates in the same order. Failing any gate does not discard the claim — it changes its status, and the status is what you read in the brief.

Gate 01
Who says this?

A named, reachable source — or the claim never entered the material.

Gate 02
About what, exactly?

Bound to your item and variant, not to a family or a lookalike.

Gate 03
Are they in a position to know?

Authority for this claim specifically, judged line by line.

Gate 04
Is it still true?

Fresh enough for a decision being made now, with the age on record.

Resulting status
All four pass → assess whether the evidence supports the claim
Missing required proof → Unresolved
Material unresolved conflict → Blocked
Machines and people

Why AI is not allowed to invent commerce facts

A language model asked whether a supplier holds a certification will produce a fluent sentence either way. That fluency is the danger: it looks identical to knowledge, and it is the mechanism by which an assumption becomes a purchase order.

So the boundary is architectural, not editorial. Machine logic may match, extract, compare, and check consistency. It may not originate a commerce fact, and it cannot promote a claim to supported. Only evidence does that.

A system that cannot say "nobody knows this" will eventually say something worse.
Machine logic does
Match a source to a subject and variantExtract stated values without rewriting themCompare sources and surface disagreementRecord age, authority, and consistencyApply the same rules to every review identically
Machine logic never
Invents a specification, price, or stock figureFills an unresolved item with an estimateResolves a conflict by choosing a sidePromotes a claim to supportedDelivers a Decision Brief without human review
Designed against

Four ways product research fails

Each principle above addresses a specific failure that can turn a reasonable decision into a loss. Naming the failure is more useful than naming the feature.

Failure 01 · wrong subject
Evidence about the neighbour

Research is assembled about a product family, then applied to one variant with a different plug, capacity, or certification. Every figure is correct and none of it is about what arrives.

Answered by principle 02
Failure 02 · borrowed authority
A source speaking outside its knowledge

A supplier states what a marketplace will accept; a reseller states what a factory produces. The statement is real, the authority behind it is not.

Answered by principle 03
Failure 03 · the tidy gap
An estimate wearing the clothes of a fact

A missing number is replaced by a category average so the report has no blanks. The blank was the finding, and now it is invisible.

Answered by principle 05
Failure 04 · silent reconciliation
A conflict resolved without you

Two sources disagree; the newer one wins; nobody is told. The one signal that should have stopped the purchase is deleted in the name of a clean answer.

Answered by principle 06

See the method applied to your decision.

The pilot is where the method meets a real opportunity. Bring one you are close to committing on.

Apply for the pilot →The product page →