Studies first: each one names its baseline, its setup, and what did not work. The systems tab holds the evaluation harnesses and products built alongside.
Result
Full-parameter SFT then DPO on Qwen3.5-2B lifted call F1 from 0.6264 to 0.7548 on a 1,277-example pre-registered internal partition and cut over-calling from 62.4% to 15.3%. The promoted model also refused all 1,319 zero-shot GSM8K questions and lost 14.9 points on 8-shot GSM8K, neither of which the promotion gate could see. One seed, reported with the regression left in.
Result
Retrieval plus LoRA SFT, DPO and on-policy distillation took gemma-4-E2B from 0.16% to 18.59% judge-scored accuracy and from 13.75% to 71.25% abstention on unanswerable probes, over a 688-probe benchmark built to catch false memories. The abstention gain is large; absolute recall stays low. Single seed, with the errata published.
Result
The hybrid beat rules alone and the model alone on every metric. Omission still caused 55 to 62% of its failures.
Exploratory
The best of four recipe changes cut omission to 52.3% of failures, the lowest in the project, and still above half.
Exploratory
An early checkpoint, before the hybrid existed: distillation cut hallucination from 0.1132 to 0.0717, with field F1 0.005 below the base model.
Prototype
A schema registry, a rules pre-pass and a distilled 1B model that beat both halves alone on every metric.
Prototype
A companion on gemma-4-E2B with tiered external memory, evaluated on a 688-probe memory benchmark.
