Dockerfiles: auto-select torch build (CUDA on x86_64, CPU on arm64); override with TORCH_VARIANT

Scorers: write numeric-only reward.json (Harbor rejects string fields); diagnostics to stdout

README: document self-contained task images (no external base image/repo/token)

Self-contained task images (FROM python:3.12-slim, own requirements.txt); remove external base-image and agent-specific coupling

README: fully agent-neutral (no agent names/specifics), generic -a <agent> run

README: drop AutoDS-specific section; add docker-less (-e singularity) run option for any agent

Reframe README: general benchmark for any Harbor agent; AutoDS specifics optional

Fix stale README/registry: single dataset, agent-neutral base instructions, data shipped on HF

Enrich TabReD base instructions with dataset context from yandex-research/tabred datasheet

Strip C1 sections leaked into TabReD base instructions (AutoDS adds them at runtime)

Add discretionary GPU hint to base instructions

Collapse to 1 folder per task; base instruction only (AutoDS C1 at runtime)

Note the 40k/10k temporal slice in registry description (transparency)

full dataset incl. real data

initial commit