Research
Anyone can publish a win. We publish the whole run: the method, the numbers with their intervals, and the exact line where the result stops being true.
- FACILITY
- SAN FRANCISCO, CA
- DIVISION
- RESEARCH
- RECORDS FILED
- 003
- LAST ENTRY
- 2026·09·30
- METHOD
- IN RECORD
- INTERVALS
- STATED
- 003 RELEASED
djev, the open source Jev, benchmarks close to it
djev rebuilds Jev's API on Google's open weights DiffusionGemma. On the comments from our Jev evaluation it ranked real defects 91% of the time to Jev's 96%, put 85% of its answers below 0.01 or above 0.99, and could not read 13% of the real defects it was sent. It fits on a 24 GB Mac mini, with nothing else running.
- 002 RELEASED
Jev answers with a number, not a sentence
Jev, TypeSafe's first model, emits no text. It scored 0.961 against the best model's 0.968 at spotting real defects, for 72 cents per ten thousand reviews against $13.08. It is the only one whose stated probabilities meant what they said, and the only one that refused outright once the request passed 32,000 tokens.
- 001 RELEASED
We put a fly's sense of smell in a security scanner
We built a fruit fly's olfactory circuit into a security scanner. It removed 97% of repeat alerts and ranked a copied function first among 2,285 candidates 93% of the time. Asked whether code was dangerous, it scored 51.1%. A coin scores 50.