Needle research
Technical reports
The Needle-1.0 report brings together the model, training experiments, results, and limitations. The underlying experiment records remain available for reference.
Reports
Supporting experiment records
Historical checkpoints, failed experiments, and detailed procedures supporting the report. These are working records, not separate model releases.
- A bidirectional marker encoder, measured against the reference API
Most models answer by writing text. Needle reads the situation and the options you give it and, in one pass, says how likely each option is. This note is the record of building that model (a…
- Iterating on synthetic data against a frozen reference panel
Before Needle could be compared with anything, we needed a test the model could not have seen and could not cheat on. This note records the first honest comparison loop: a frozen panel of 1,000…
- Composition curriculum and the swapped-choices failure
The previous model looked like it understood policies and turned out to be reading the position of the answer. This note is the slow, unglamorous work of teaching the small skills separately…