9.4× faster and 90% lower cost, with 10.9% higher verdict accuracy.

Claims research, built for the workflow

A claims research agent checks a real-world claim against the evidence store attached to that claim. It searches, reads the passages it retrieves, decides whether more context is needed, and returns one of four verdicts: Supported, Refuted, Not Enough Evidence, or Conflicting Evidence.

We ran the comparison on the public AVeriTeC benchmark's 500 held-out development claims using the same search tool, retriever, prompt scaffold, greedy decoding, and eight-turn limit. This is a controlled benchmark experiment, not a customer deployment.

Training a model for claims research

We trained Qwen3.5-4B with GRPO to maximize exact-match verdict accuracy, using a LoRA adapter and complete search-and-reasoning trajectories.

Training used complete search-and-reasoning trajectories rather than treating each turn as a separate record. That preserved the sequence the model actually sampled and kept long episodes from receiving extra weight simply because they contained more turns.

Higher accuracy with faster, lower-cost processing

The trained Qwen3.5-4B model delivered a 10.9% relative improvement in verdict accuracy over Claude Haiku 4.5. The paired confidence interval was +4.5% to +17.3% relative, and the official AVeriTeC score improved 11.3% relative.

The workflow also used 44% fewer searches, with faster processing and lower cost across the claims run.

A model tuned to the claims workflow

The trained model learned the full claims loop: finding relevant evidence, deciding when it has enough, and returning the verdict. The result is a higher verdict score with nearly half the searches of Haiku 4.5.

For teams building claims and fact-checking agents, the experiment gives a concrete benchmark for training on the workflow your model actually runs, then measuring the result on held-out claims.

Routing between three models

A trained small model does not have to be the whole deployment. We deploy the 4B as the default lane and put a lightweight router in front of it. The router reads the claim before any model runs: it embeds the claim text with text-embedding-3-large, adds the claim's metadata (country, reporting source, claim year, and whether a speaker is named), and feeds both into a regularized logistic classifier. Based on that score, each claim goes to the trained 4B, to DeepSeek v4-flash, or to Claude Opus 5. The decision itself adds no model calls and almost no latency.

The router is trained on 2,896 separate training claims, and every number here comes from the 489 held-out claims it never saw. It turns the three models into a single cost dial. The trained 4B alone scores 71.1% at $1.75 per 1,000 claims, above DeepSeek v4-flash at 67.7%. A mid-dial setting sends 76% of claims to the 4B and 24% to Opus: 73.6% accuracy at $27 per 1,000 claims, which is 95% of always-Opus accuracy at under a third of its cost. The top of the dial is always-Opus at 77.5% and $89.

The router shown here is a deliberately simple example: an off-the-shelf embedding, a handful of metadata fields, and a logistic classifier. Its difficulty signal is modest, so at this stage most of the value per dollar comes from blending strong endpoints, and there is clear room to grow the routing signal itself. What makes the cheap end of the dial worth standing on is the trained model, which raised that floor from 57% to 71% on this benchmark.