AI transparency report

Show exactly how AI supported your screening

Generate methods text, review the recorded model configuration, measure human-AI agreement, and export screening decisions with the AI metadata available in your project.

Human decisions remain finalAI ranks and advises.
Configuration documentedModels and parameters are named.
Exportable evidenceMethods text and audit CSV.

Methods reporting

Turn project settings into a methods paragraph

Lumina generates editable methods text from the configuration recorded for your project. It identifies the embedding model, TF-IDF classifier, retraining interval, relevance-feedback settings, optional AI consultations, stopping rule, and final screening counts when those data are available.

The paragraph is a reporting aid, not a certification. Review it against your protocol, journal requirements, and the workflow your team actually followed before publication.

Generated methods text Editable
Title-abstract screening was performed using Lumina. Papers were initially ranked by semantic similarity using OpenAI text-embedding-3-small embeddings. After sufficient screening decisions, TF-IDF with Multinomial Naive Bayes contributed to queue prioritization and was periodically retrained. All final screening decisions were made by human reviewers.

Recorded configuration

See the components behind the ranking

The report exposes the configuration Lumina uses instead of replacing it with a generic “AI-powered” label.

Embedding model
text-embedding-3-small
Classifier
TF-IDF + MultinomialNB
Relevance feedback
Rocchio α 1.0 · β 0.5 · γ 0.25
Human control
AI recommendations are advisory

Audit CSV

Export decisions and available AI metadata

The audit export contains citation fields, screening order, reviewer decisions, exclusion reasons, notes, current paper scores, the project model version at export, and Pixel-Bot recommendation, confidence, and reasoning when the assistant was consulted.

  • 01Connect human decisions to reviewers and timestamps.
  • 02Inspect agreement between recorded AI recommendations and final human labels.
  • 03Retain exclusion reasons, notes, and citation identifiers in one file.
project_audit.csv · selected columns
paper_id,human_decision,ai_recommendation,ai_confidence
1492,include,include,0.91
1493,exclude,exclude,0.84
1495,exclude,include,0.62

Know what the report does not prove

A transparency report improves documentation, but it does not validate a stopping decision, guarantee recall, establish compliance, or prove that a review is methodologically rigorous. Those conclusions depend on the protocol, reviewer workflow, validation approach, and reporting requirements.

The audit CSV records the project model version and current paper scores when exported. It is not a point-in-time snapshot of every ranking score for every historic decision.

Reporting references