Research note · 6 September 2026
Strong reasoning.
Fast reconstruction.
A first test of dividing answer generation between Astra and Spark.
What we tested
Six samples ran concurrently: two direct Astra high outputs, two direct Spark low outputs, and two sequential Astra high → Spark low pipelines. Every sample received the same Python regex-engine task and was evaluated by the same independent checker.
For the pipeline, Astra produced a hierarchical implementation plan with executable parser and matching helpers. Spark was instructed to preserve those helpers and implement the surrounding public APIs. Plans had a ceiling of 1,800 tokens, measured with the same tokenizer across both stages. No model was fine-tuned.
Results
| Path | Correct | Generation E2E | Final tok/s, E2E |
|---|---|---|---|
| Astra high | 2/2 | 128.19 s | 14.8 |
| Spark low | 0/2 | 28.27 s | 89.6 |
| Astra → Spark | 2/2 | 91.97 s | 19.4 |
Both pipeline candidates passed all 82 check groups. Both preserved Astra’s core functions by syntax-tree comparison. Plans contained 1,599 and 1,626 tokens. Pipeline generation times were 84.85 and 99.08 seconds; direct Astra times were 144.51 and 111.87 seconds.
What the result means
Pipeline median generation time was 28.3% lower than Astra’s median in this batch. This supports investigating the division of reasoning and expansion, but two runs on one task do not establish general reliability or a stable speed advantage. Final output was only about 1.1 times the plan length, so the experiment does not demonstrate substantial learned compression.
Generation E2E includes startup, model calls and cleanup, but excludes independent test execution. Median generation time plus each candidate’s test execution was 128.84 seconds for Astra, 29.06 seconds for Spark and 93.04 seconds for the pipeline. Shared service capacity can affect latency. Passing finite tests does not prove correctness for all inputs.
What comes next
Train an answer encoder and a fast decoder to discover shorter representations that preserve required behavior. Separately train a prompt controller to help a frozen frontier model emit compatible representations. Test that connection early: a representation that reconstructs well may still be hard for the frontier model to produce.
Success means preserved quality and lower total completion time on unseen tasks, including controller and decoding overhead. The learned codec and controller remain research work.