BrainHack 2026 TIL-26 Workspace

This repository contains the local development scaffold for the TIL-26 task containers:

• ae/ - autonomous exploration, with til_environment available through the til-26-ae submodule.
• asr/ - speech recognition.
• cv/ - object detection.
• nlp/ - RAG question answering.
• noise/ - adversarial noising for CV inputs.

The source scaffold is mirrored from /Users/n00bcak/programming/til-26, excluding the dormant/private til-26-finals submodule.

Python And uv

Each task folder owns its own uv project and virtual environment. Do not create the venvs as part of repository setup; create them only when working in that task.

cd ae
uv python install 3.10
uv sync --locked --group dev

Repeat per task folder as needed. pyproject.toml is the source of truth for direct dependencies. requirements.txt is generated only for Docker installs.

Requirements Drift

Generate Docker requirements only on Linux:

scripts/sync-requirements
scripts/check-requirements

The sync command runs this in each task folder:

uv export --no-hashes --no-emit-project --no-dev -o requirements.txt

On macOS, scripts/check-requirements verifies structure and intentionally skips regeneration because Linux CI is the canonical export environment.

NGC Container Constraints

ML task Dockerfiles use nvcr.io/nvidia/pytorch:25.11-py3 and install with the container constraint file:

pip install --no-cache-dir -r requirements.txt -c /etc/pip/constraint.txt

Before adding ML dependencies, capture the actual container package truth:

scripts/capture-ngc-constraints

That records python -VV, pip freeze, and /etc/pip/constraint.txt under constraints/. Validate new ML dependencies inside the NGC base container before locking them.

Build Tags

Workbench builds use til build. Laptop builds should use local tags such as:

docker build -t til-local-ae:dev ae

Use TEAM_ID tags only for competition submissions.