Evidence
Public evidence
SentX publishes the records behind its demonstrations, including prompts, answers, errors and limits. They are first-party observations, not independent benchmarks.
Updated
Records
Reading a complete research paper
- What was tested
- One request supplied the extracted text of a 25-page research preprint and asked for an evidence brief. Browser upload was not tested.
- Source paper
- “Recalling Too Well: Sycophancy Evaluation and Mitigation in Memory-Augmented Models” by Shelly Bensal, Axel Magnuson, Aparna Balagopalan and Daniel M. Bikel (arXiv:2606.10949v2, CC BY 4.0). SentX was not involved in this research, and the paper does not evaluate SentX.
- Result
- A reviewer checked the answer against the paper in 11 source checks: 5 supported, 4 supported with qualifications or gaps, 2 not supported.
- Recorded errors
- Two substantive errors: a claim that every comparison showed an increase, which one table contradicts, and an unsupported reading of a percentage in the abstract. Three further qualifications or omissions are listed.
- Limits
- One paper, one answer and one first-party review. It does not measure general accuracy, retained memory, PDF upload or speed.
Continuing a project from a supplied checkpoint
- What was tested
- Two requests in separate sessions used an invented project. The second request included the full checkpoint written in the first.
- Result
- Both answers kept the supplied facts. The second updated the inventory to 70 seed packets and the budget to AED 2,700 (2,400 + 300).
- Recorded errors
- Three recorded issues: an unsupported checkpoint date, an inferred receipt date, and supplied facts described more strongly than given.
- Limits
- One synthetic project and two answers. The checkpoint was pasted in full, so this does not test memory across sessions.
How to read these records
- Each JSON file contains the prompts, final answers, timestamps in Dubai time (GMT+4), hashes and the reviewer's findings.
- Internal reasoning is excluded.
- The model label is the public product label; the underlying model version was not independently verified.
- SentX prepared these records. They are not independent endorsements.