Plicara is Adrian Tame’s independent AI research lab. We investigate how AI systems behave, how to measure them, and which changes make them more reliable. The work takes the form of reproducible studies, benchmarks and practical tools.
Plicara is from Latin plicare, to fold.
Selected public entry points are listed below. Each repository owns its project metadata and reproduction instructions; this profile is an introduction, not a project-status board.
| Repository | Ships |
|---|---|
regexbench |
regexbench on PyPI — regex correctness and ReDoS evaluation |
labloop |
labloop on PyPI — keep a change only if it measurably helps |
pi-audit-loop |
@plicara/pi-audit-loop on npm — a one-pass review, change, and verification workflow for the pi coding agent |
regexeval-2026 |
The eleven-model regex run: runner, predictions, scored results, methodology |
articles |
The code behind published research: the scripts, the results, and the exports every figure is built from |
plicara-brand |
The brand as code: tokens, marks, logo set, and the generators that draw them |
plicara.github.io |
plicara.ai — the site and the articles |
Research explains the questions and findings. Tools provides installation examples. Benchmarks collects the evaluations and their methodology.
Evidence you can inspect
- Code, methods and recorded outputs belong beside the claim. Models are one possible research artifact.
Reproducible research
- We publish methods and evidence where available, and state which results can be replayed and which still have evidence gaps.
Reliable systems
- Boring infrastructure is a feature, and we take care to build systems that are robust and generally scalable.
Simple is not the enemy of powerful
- We reach for the plainest thing that works, and our models, research, and benchmarks carry this ethos.
Code is Apache-2.0 unless a repository's LICENSE says otherwise (pi-audit-loop is MIT, benchmarks data is CC BY-NC 4.0).
Model weights are licensed per release, with the decision recorded on the model card.