Open questions, published either way.
The products depend on problems nobody has fully solved: attribution, retrieval over versioned text, scheduling against real memory. We write up what we find, including the negative results.
All research
- Benchmark
Spacing intervals at topic level, not card level
An in-progress evaluation of scheduling review by topic mastery rather than per-item, using StudyFlow's early cohort data.
- In progress
- learning
- scheduling
- evaluation
- Note
Attribution is not accuracy
Grounded generation systems are routinely evaluated as if faithfulness to retrieved text were the same thing as being right. It is not. We separate the two and show where the confusion costs you.
- evaluation
- retrieval
- grounding
- Experiment
Latency budgets for interfaces that feel instant
A practical set of per-interaction time budgets, and what we give up to hold them. Measured on the products in this lab rather than borrowed from a 1968 paper.
- performance
- interface
- engineering
- Paper
Point-in-time retrieval over amended statutes
Legal text is versioned, and most retrieval systems flatten it. We describe the indexing scheme behind NexLex's as-of queries and the failure cases it removes.
- Preprint
- retrieval
- legal
- systems