zeroshot research blog

What we think, and what we tested.

  1. What the next run remembers

    How zeroshot turns completed work into useful lessons and recalls them through file and keyword indexing when an agent needs them.

  2. Fast workflows with code and models in zeroshot

    A retrieval case study of fast zeroshot workflows that combine deterministic code and decision models, with measured ranking quality, latency and cost.

  3. The SDLC, reimagined for the post-review era

    How teams can use zeroshot to build, test and review changes while people spend more time deciding what to build.

  4. A dark factory for time series collection

    How a zeroshot graph can discover and validate original public time series for training and evaluating time series forecasting models.

  5. A small model with a zeroshot loop matches a larger one at 58% of the cost

    On 20 ProgramBench tasks, GPT-5.6 Luna in a build-and-check loop reached GPT-5.6 Sol's single-shot score for a median 58% of Sol's cost, and at Sol's cost scored higher on 15 of 20.