The Fieldbook evidence model

Most bad commerce decisions begin with a labeling error: an estimate is presented as a result, a platform claim is treated as accounting truth, or a correlation becomes a root cause.

Observed

An observed fact came from a named source: a rendered product page, network response, catalog record, merchant diagnostic, transaction ledger, browser test, or timestamped AI answer.

Record enough context to reproduce it: source, time, product or journey, market, device, and test conditions.

Attributed

Advertising and analytics platforms assign credit according to their own identity rules, attribution windows, deduplication logic, and modeled conversions.

Modeled

Forecasts, opportunity estimates, probabilities, and expected impact are models. Show their assumptions and confidence range. Never rewrite them as measured results.

Hypothesis

A useful hypothesis explains an observation and names a test. It should identify what will change and what result would strengthen or weaken the explanation.

Verified

A result is verified when a defined change is followed by a retest. Verification can show that markup now matches the page, a feed now carries the right price, an event is no longer duplicated, or a monitored answer changed.

That retest may confirm the fix without proving revenue impact. Commercial effects often need a longer measurement period and a separate design.

Minimum record

• Product, page, event, or journey
• Source and collection method
• Time, market, and relevant test conditions
• Observed value
• Expected value or acceptance rule
• Confidence and open questions
• Owner and next check