Ponys.ai Research Evidence Library

Reproducible AI character evaluation resources

AI Companion Memory Retention Test Set

Use this open test instrument to compare systems without hiding failures, sample sizes, model versions, or conflicts of interest.

Protocol coverage

Test case coverage by locale

This release contains 140 preregistered cases, including 20 for English. Run the preregistered cases, preserve the denominator, publish negative outcomes, and cite the version used.

Measured dimensions

Scores use a 0-2 evidence rubric. A zero indicates missing, contradicted or unsafe evidence; one indicates partial evidence; two requires complete reproducible evidence.

Download and cite

Download test-cases.csv · CITATION.cff · BibTeX · JSON Schema

This is an official Ponys.ai test instrument, not independent editorial research. Results begin as not_collected; publishers must disclose methods, failures and conflicts before reporting scores.

Audited research targets

These indexable product paths are included for reproducible test setup and deep-link attribution.