Butanium

Write up research experiments and findings as an interactive, self-contained HTML report published as a claude.ai Artifact, with figures and data exploration. Use when creating a research report from experiment results.

How to write a comprehensive research report for your supervisor summarizing experiment findings, hypothesis tests, and data exploration. Use when you need to communicate results from multiple experiments with proper analysis, visualizations, and interactive data exploration.

How to send notifications to the human supervisor via ntfy.sh. Use when you need input, hit a blocker, or update them on your progress.

Cost and latency optimization for Anthropic API usage. Covers prompt caching, batch API, and when to combine them.

Standard experiment folder structure and templates. Reference for creating or validating experiment folders.

Read model outputs, samples, or log entries by eye and report specific observations about what's in them. Use when asked for a "qualitative" read/analysis, or when the goal is to understand WHAT is in the data rather than count patterns. Not for numerical summaries, metric computation, or classifier-driven sweeps.

How to evaluate research samples using structured JSON output from claude -p. Covers criteria writing, the core judging pattern, and practical examples.

Core research principles for hypothesis-driven investigation. Shared across orchestrator, scientists, and colleagues.

Browse experiment outputs (JSONL/CSV/inspect-ai .eval) together with the human via samplescope — start the server, then drive the human's browser view from the terminal with `sscope view` (open files, jump to rows, filter, mark, judge, pin plots). Use when the human wants to look at samples/results/transcripts together, asks to "open X in the viewer", or when you want to show them specific rows instead of pasting walls of text.

Browse experiment outputs (JSONL/CSV/inspect-ai .eval) together with the human via samplescope — start the server, then drive the human's browser view from the terminal with the `viewer` CLI (open files, jump to rows, filter, mark, judge, pin plots). Use when the human wants to look at samples/results/transcripts together, asks to "open X in the viewer", or when you want to show them specific rows instead of pasting walls of text.

Persistent Python REPL server that avoids reloading heavy packages (torch, transformers, nnterp, vllm). Use when running multiple Python commands with heavy imports, loading ML models, or doing iterative testing where variables should persist between executions.

Chat with / sample from Tinker training-run checkpoints together with the human via tinkerscope — start the server, then drive the human's browser playground from the terminal with the `tinkpg` CLI (select a run, stream a chat, draw an n-sample distribution, compare two runs). Use when the human wants to poke at a fine-tuned checkpoint's behavior, browse discovered Tinker runs, see what a model "usually says" to a prompt, or A/B two runs — and when you want them to watch the samples land in their browser instead of pasting walls of text.

Find agent transcripts from a completed Nomic game