Explorer

Search 2,730 indexed units across two corpora. C1 is core Python (stdlib, pandas, polars, pydantic, FastAPI, typer, httpx, pytest, ruff, mypy, uv); C2 is AI tooling (anthropic, openai, instructor, tenacity, MCP, a local eval harness). Source is read live with inspect.getsource in the corpus interpreter — it is the real file, not a copy someone typed. Live run executes whatever is in the box.

Rosetta

200 SAS→Python and 100 R→Python mappings, frozen to JSON. Run existence checker re-resolves every Python symbol referenced by every entry, in the corpus interpreter, right now. Symbols are extracted from the entry's own code by AST, so an author cannot claim a mapping teaches an API without that API appearing in code the checker resolves.

Creator

Type an intent, get a typed package with a CLI, a FastAPI surface, tests and a Dockerfile. The log streams live. Failures are fed back to the model verbatim and it retries; after the attempt budget it reports FAIL.

Eval Lab

Edit the suite JSON, run it. Two full passes over the model: baseline, then a degraded system prompt. Read the degraded verdict first.

Comparator

Unified code diff plus a behaviour diff obtained by running both programs sandboxed and diffing real stdout, with wall time and peak memory for each side.

Documentor

Builds the zip and records a signature. The download is the same bytes a reviewer gets.

Learn Path

Track A is Python to production and is built on PyLab's authored curriculum cards, imported rather than rewritten. Track B is AI tool development and is new.

Feedback

Files a dated, checkboxed dev task to state/dev_tasks.md. It is a file on this box, not a ticket in someone else's system.

Written by a model, every run

Creator's package source: the README, the library module, the CLI, the FastAPI surface and the tests. These differ run to run and can be wrong — that is exactly why nothing ships until the HARNESS is green.

Eval Lab's answers: every case is a real API call. Scores move between runs. A single run is a draw, not a result.

Templated — identical every run, no model involved

pyproject.toml, the Dockerfile and py.typed are generated from a fixed template in pyforge/scaffold.py. The model is explicitly forbidden from writing them and its versions are discarded.

The Rosetta corpus is curated data with an automated checker, not runtime model output. The degraded system prompt in the Eval Lab defaults to a fixed string.

Deterministic, no model at all

Explorer's index and source, the sandbox, the harness, the comparator's diffs and the bundle's SBOM and hashes. If any of those is wrong it is a bug, not a sampling artefact.

What the sandbox does and does not do

Code you type in Live run and Comparator executes as an unprivileged user in an empty network namespace with CPU, memory, process and file-size caps, in a throwaway directory, with a whitelisted environment — the server's own environment is never passed through. It is defence in depth, not a jail: the child can still read world-readable files. Every secret on this box is 0600 root, which that user cannot open.

Creator output is different. The harness runs generated code as root in the tool venv, because building a wheel and a Docker image needs it. That is a deliberate, bounded exposure: the code came from a model under a fixed system prompt, not from a stranger, and the studio is auth-gated.

Cost and latency figures

Token counts are estimated from character length and prices are per-model constants. They are an order-of-magnitude guide for engineering decisions, not a bill.

Coverage

Coverage is statement coverage of the generated package only, threshold 80%. It says the lines ran, not that the assertions were meaningful.

recent tasks