Personalization accuracy study: study protocol

Define the decision

This page is a proposed research protocol, not a completed study or a report of findings. The decision is: Are generated opening claims supported by the source available at generation time? The observation unit is one factual claim, nested within an opening line. Define the population and owner before collecting records; do not substitute an available convenience dataset without documenting the change.

Work through the procedure

1. Store the exact source excerpt, retrieval date, prompt version and generated claim. Separate facts from stylistic suggestions.
2. Review claims against the saved source without seeing the preferred model name. Label supported, contradicted, unsupported or unverifiable.
3. Before collection, write the primary outcome, observation window, exclusion rules and stopping conditions. Preserve excluded observations with a reason rather than quietly removing them.
4. Pilot the procedure with fictional or owned test data, resolve ambiguous fields, and freeze a dated protocol version before the main run.

Worked example

The following is a synthetic example for this procedure, not a customer result or performance benchmark.

claim: opened a Berlin office; source: recruiting a remote employee in Germany; label unsupported
Suggested record fields: observation_id, condition, evidence_reference, outcome, exclusion_reason, reviewer, protocol_version
Status: illustrative record only; no study has been run for this page.

Read the result

Report claim accuracy and message-level incidence separately. Several claims in one message are correlated observations. Keep the numerator, denominator and missing evidence visible. If the available observations cannot answer the registered question, report that limitation rather than selecting a more favorable metric after collection.

Collect the evidence

Observation unit: factual-claim.

Saved source excerpts, generation metadata and claim-level outputs.

Hide model name during review; retain unverifiable claims separately.

Analysis: Supported, contradicted, unsupported and unverifiable claim counts; also aggregate by message.

Download the JSON collection template under Source references. Set an owner, eligibility rules, outcome definition, observation window and sample justification before collecting records. Templates contain no participant data or results; keep original private records outside the public website.

Check before moving on

1. Name the person responsible for collection and review.
2. Check that the evidence can be inspected without exposing private messages or credentials.
3. Record deviations from the protocol and analyze their possible effect.
4. Retain a dated, redacted evidence worksheet with the final interpretation.

Limits and next action

A fluent paraphrase is not evidence. Exclude inaccessible sources from verified counts and show them as unresolved. Publish results only after the evidence, method and limitations have been reviewed. This protocol provides no benchmark, expected lift or completed-study claim.

Source: Method or workflow reference