Bridgewater’s text-to-SQL model beat frontier systems—but on a repaired benchmark, not the BIRD leaderboard

Thinking Machines Lab has publicly identified Bridgewater as the enterprise user behind ReViSQL-K2.6, a custom text-to-SQL model trained with its Tinker post-training platform. In an October 15 anniversary post, the company said the model beat Claude Fable 5 and GPT-5.6 Sol on accuracy at 15% of their inference cost. The underlying technical report, code and cleaned data are public.

The headline result is 92.97% accuracy when ReViSQL-K2.6 generates 16 candidates and selects by majority execution result. Thinking Machines reports a cost of $0.56 per task. With greedy single-sample decoding, the model scores 91.37% at $0.035 per task. Those are unusually strong cost-and-accuracy numbers for a system without a multi-stage agent scaffold.

The benchmark label matters

This is not a new score on BIRD’s official leaderboard. The evaluation uses Arcwise-Plat-SQL, an expert-corrected form of BIRD Mini-Dev. The researchers say earlier Arcwise work corrected errors in 32.3% of instances and their second pass brought the detected error rate to 52.8%. The 92.96% “human” line is therefore a proxy used on that repaired set, not evidence that the system has surpassed people across production databases.

That distinction does not make the result unimportant. It changes what the result proves: ReViSQL-K2.6 performs exceptionally on a cleaned, inspectable evaluation whose labels the authors argue are more reliable than the original benchmark.

Data quality supplied most of the lift

The team audited 2,500 BIRD training examples and reports that 61.1% contained at least one problem; 52.1% had incorrect gold SQL, 26.2% had flawed questions and 18.2% had wrong external knowledge. Categories overlap. Training Kimi-K2.6 on the released BIRD-Platinum set produced 88.55% accuracy before the final reward changes.

The recipe then addressed two reward failures. A bounded SQL-equivalence verifier found that 32.8% of positive execution-match rewards in a pilot were not fully equivalent to the gold query. The team downweighted those false positives. It also required explicit requirement and verification blocks because 24.2% of pilot validation failures came from ignoring supplied external knowledge.

For enterprise teams, the practical lesson is narrower than “fine-tuning beats agents.” Before adding more orchestration, audit whether the training labels and reward function reinforce the behavior the business actually wants. Bridgewater’s disclosed result suggests that cleaning the ruler can matter as much as changing the model—but deployment evidence on live schemas is still needed before treating a repaired benchmark as production proof.