Results
What we measured, and how.
Each result says how we measured it and what data we used.
DiaOct 25, 2026
24 / 100Readiness of a sales CRM before any modelling work
The readiness score tells you before anyone asks
Each database connection gets a readiness score from 0 to 100, calculated without AI. A sample CRM scored 24 of 100 with 3 of 152 tables approved. That is the point. You see what Dia can answer for before the business does.
Dia pull request #120Method and dataDiaOct 24, 2026
345Business entities found in 342 tables
Dia proposed 345 business entities from 342 tables in 1.1 seconds
Semantica turned a 342-table database into 345 business entities and 788 possible relationships, in about a second. The planner uses only relationships a person approved. None are used until someone does.
Dia pull request #117Method and dataForgeLabOct 19, 2026
289 of 379Questions where our tuned model came out ahead
Our tuned 14B model beat Claude Opus 4.8 on 289 of 379 Oracle questions
We trained Qwen3-14B on public Oracle documentation and put it head to head with Claude Opus 4.8 on 379 questions neither had seen. Ours came out ahead on 289 of them. Opus won 53, and 37 were ties.
Oracle evaluation, run 11Method and dataDiaOct 17, 2026
3 of 3Hidden write commands blocked in the test (stacked, direct, nested)
Three hidden writes, three refusals
Every query Dia's chat writes must be a single read. Code refuses anything that writes, even a DELETE hidden inside a larger query. Each query also runs in a read-only transaction that is always rolled back. 75 automated cases cover the rest.
Dia pull request #104Method and data