Public project · Interactive demonstration

Backend Study Lab

Make an architecture decision from evidence, then inspect a review rubric.

Choose a scenario

Synthetic fixtures. This browser demo runs locally without an LLM or network service.

Evidence

Decision exercise

Your reasoning

Inspect the workflow

Researcher and challenger notes

Reference experiment and review rubric

Public repository

View source on GitHub · Read the architecture

Also included: a deterministic concurrency simulation, recorded results, and four simulation checks. Read the experiment.

Source and evaluation

Download runnable Python source

python3 study_lab.py run --scenario worker-contention
python3 study_lab.py evaluate

The source includes three fixtures, seven contract checks, and a model-adapter interface. Output must cite known evidence IDs and include an experiment, measurements, uncertainty, and a rollback condition.

Keyword coverage can miss synonyms and accept weak answers. It is not a correctness score. Contract validation cannot establish semantic support for a citation. Human review remains necessary.

Offline mode has no model-token cost. Elapsed time is measured by the CLI; it is not a model or production benchmark.

What this demonstrates

The harness makes evidence and review boundaries explicit. A researcher proposes hypotheses, a challenger identifies missing information, and a reviewer checks a small contract before the exercise is accepted. In the baseline these stages are represented by fixture outputs; a model adapter can replace that provider.