name	scientific-verification
description	Verify ML correctness, library usage, formatting, tests, dead code, and client/backend integration.
argument-hint	[scope]
allowed-tools	Read, Write, Glob, Bash
context	repo
agent	three-mlagents-evaluation-scientist

Scientific Verification Skill

Use this skill before declaring work complete.

Backend Checks

Run the broad backend checks unless the change is documentation-only:

uvx ruff check backend/examples backend/mlagents backend/main.py backend/services backend/tests
uvx ruff format --check backend/examples backend/mlagents backend/main.py backend/services backend/tests
uv run --project backend python -m unittest discover -s backend/tests
uv run --project backend three-mlagents list --trainable-only

For a changed trainable task, also inspect the task and run a short fixed-seed train/evaluate sanity pass. If full benchmark training is too expensive, state that clearly and do not claim final performance.

Frontend Checks

Run:

cd client && pnpm run build

For route or WebSocket changes, verify with live backend and frontend servers. Confirm that trainable tasks call real backend routes and roadmap tasks do not show fake training controls.

Scientific Review

Check:

Gymnasium API compliance and dtype/bounds correctness.
SB3 algorithm compatibility with observation and action spaces.
seed handling and reproducibility.
termination versus truncation semantics.
reward signs, units, and scaling.
episode metrics and evaluation protocol.
task registry metadata, research tier, and publication role.
optional dependency failures are graceful.
frontend claims match backend capability.

Dead-Code And Artifact Review

Search for:

rg "three_mlagents|backend/three_mlagents|onnxruntime|hand-rolled|fake policy|TODO" .
rg "console\\.log|debugger" client/src
git status --short

Do not remove unrelated user changes. Do not commit generated policies, caches, build output, TensorBoard runs, or dependency directories.

Completion Standard

Smoke tests can support the evidence, but they do not satisfy a full implementation request by themselves. Completion requires code, docs, tests, and client/backend behavior to agree.

Scientific Verification Skill

Use this skill before declaring work complete.

Backend Checks

Run the broad backend checks unless the change is documentation-only:

uvx ruff check backend/examples backend/mlagents backend/main.py backend/services backend/tests uvx ruff format --check backend/examples backend/mlagents backend/main.py backend/services backend/tests uv run --project backend python -m unittest discover -s backend/tests uv run --project backend three-mlagents list --trainable-only

Frontend Checks

Run:

cd client && pnpm run build

For route or WebSocket changes, verify with live backend and frontend servers. Confirm that trainable tasks call real backend routes and roadmap tasks do not show fake training controls.

Scientific Review

Check:

Gymnasium API compliance and dtype/bounds correctness.

SB3 algorithm compatibility with observation and action spaces.

seed handling and reproducibility.

termination versus truncation semantics.

reward signs, units, and scaling.

episode metrics and evaluation protocol.

task registry metadata, research tier, and publication role.

optional dependency failures are graceful.

frontend claims match backend capability.

Dead-Code And Artifact Review

Search for:

Do not remove unrelated user changes. Do not commit generated policies, caches, build output, TensorBoard runs, or dependency directories.

Completion Standard

Smoke tests can support the evidence, but they do not satisfy a full implementation request by themselves. Completion requires code, docs, tests, and client/backend behavior to agree.

scientific-verification