Skip to content
coilycoPublic

About

Human-graded behavior evaluations: the grader, for any consumer.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

housecast

Human-graded behavior evaluations, for any agent

housecast // grade - Human-graded behavior evaluations, for any agent

housecast is the grading half of an evaluation: the case and dataset schema, the pairing rule, one-keystroke annotation, the grading page, the room-facing deck, and the MCP tool-description loop. It knows nothing about who the subject is. A consumer brings its own cases, entity projection, profile, and runner, and housecast grades what came back. It ships no profile, so every board names its own.

docs/FEATURES.md is the inventory. The name housecast is held on PyPI as a 0.0.1 placeholder. The release train that turns a housecast-v* tag into a real upload is wired, and until the first tag runs it, consumers install from GitHub.

The name

A house cast is the resident standing company of a theatre: a fixed troupe held by one house under one set of conventions, able to mount any role in its book. house is the authority noun. cast carries both casting a play and casting metal, and that second sense is the immutable bundle.

Two other readings of the string are live, and neither one is this project.

  • house plus forecast - housecast also parses that way, and three of the four existing GitHub repositories under the string are dead housing-price forecasters. This engine forecasts nothing.
  • type casting - cast is a live Python concept, so a Python reader lands on coercion for a beat before they land on the theatre. This engine does no type casting worth the name.

housecast unhyphenated is the repository, the import name, and the primary distribution. house-cast is held defensively and never shipped.

Quick start

git clone https://github.com/coilyco/housecast.git
cd housecast
uv sync --extra room
export ROOM_PROXY=http://localhost:11434   # your endpoint's base URL, no /v1
export ROOM_MODEL=llama3.2                 # a model that endpoint serves
uv run --extra room housecast room serve \
  --subjects examples/two-agents/subjects.json --log /tmp/room.jsonl

That serves the live room on port 8767 with two example agents. Set the two variables first: without them the room calls the author's own endpoint, and every answer shows as failed. A hosted endpoint also needs ROOM_PROXY_KEY. docs/room.md covers the endpoint, key, and model settings, and examples/ has the steps. just with no arguments lists the dev verbs, and just check is the offline gate: lint, format check, types, and tests in one recipe.

Installing it elsewhere

No release is on PyPI yet, only the reserved name. Depend on it from GitHub with uv:

[project]
dependencies = ["housecast"]

[tool.uv.sources]
housecast = { git = "https://github.com/coilyco/housecast.git", tag = "housecast-v0.1.0" }

That is the same shape the estate already uses for aos-eval. Add the eval extra when the consumer needs the grader rather than the base package alone. Once a housecast-v* tag has run the train in docs/publishing.md, pip install housecast is the shorter path.

Layout

  • housecast/ - the package: digest.py, grade/, mcpeval/, mcp/.
  • Evaluations live in the repo that consumes their result, never here.
  • scripts/ - release, kit sync, and the MCP demo recorder.

License

MIT. Kai Ase Siren holds the copyright. See LICENSE.

See also

About

Human-graded behavior evaluations: the grader, for any consumer.

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages