param-decomp CLAUDE.md

param-decomp CLAUDE.md is an instructions file for coding agents from goodfire-ai/param-decomp. It costs 504 tokens per session, scanned A, original, MIT.

Repository instructions for param-decomp, a project that studies how model parameters can be decomposed and analyzed. They define which guides, configuration files, specifications, and local instructions to read for different tasks.

In plain words
What is it for?
Use them when planning experiments, choosing targets and objectives, running parameter sweeps, validating results, or changing the project's core, experiment, training, or clustering code.
Why use it?
They reduce the risk of running experiments incorrectly or breaking rules about portability, evidence, numerical behavior, and project structure.

Instructions file

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add instructions/goodfire-ai/param-decomp/claude-md
Clone the repo
git clone --depth 1 https://github.com/goodfire-ai/param-decomp

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for param-decomp CLAUDE.md

README.md
[![agentmods](https://agentmods.dev/badge/instructions/goodfire-ai/param-decomp/claude-md.svg)](https://agentmods.dev/instructions/goodfire-ai/param-decomp/claude-md)
Your own site
<a href="https://agentmods.dev/instructions/goodfire-ai/param-decomp/claude-md"><img src="https://agentmods.dev/badge/instructions/goodfire-ai/param-decomp/claude-md.svg" alt="Measured on agentmods" height="20"></a>
Per session 504 This file is loaded in full into every session.
When invoked 504 The same file — it is already loaded in full.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00504 $0.00504
Opus 5 $0.00252 $0.00252
Sonnet 5 $0.00101 $0.00101
Haiku 4.5 $0.00050 $0.00050

Measured 3d ago against content hash bdce27798360, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

param-decomp CLAUDE.md scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

CLAUDE.md · 35 lines

How it starts

The opening of the file, as written. The whole thing — 35 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Repository guidance

Before planning, running, or interpreting a decomposition experiment, read both guides:

  • docs/handbook.md — the science, evidence standards, failure modes, and interpretation limits.
  • docs/skill.md — the repository-specific recipe for targets, objectives, sweeps, validation, and analysis.

For other work, start with the smallest relevant source of truth:

  • README.md — installation, runnable entry points, datasets, packaging, and development commands.
  • CONFIGS.md — which configurations belong in the repository and how they stay portable.
  • param_decomp/core/SPEC.md — the trainer's normative numerical contract.
  • The nearest module-level CLAUDE.md — local architecture and interfaces. These exist under core, experiments, experiments/lm/pretrain, and clustering; targets carries a README.md.

Repository-wide constraints

  • param_decomp/ must not know where it runs: no scheduler, submission, code-shipping, cluster path, mount, partition, or team namespace. Paths are explicit required inputs, and configs identify external resources by portable names. Deployment adapters may depend on the library; the library may never depend on them.
  • Current training is JAX. The retired Torch implementation is only a semantic oracle at git tag torch-oracle; nano_param_decomp/ is a standalone Torch reference and is not imported by either package.
  • Keep the functional core pure and put I/O at entry points. Encode invariants in types when possible and assert the rest. Fail closed rather than adding fallbacks, compatibility shims, or degraded modes.
  • Import public names from the modules that define them; package-level re-exports are exceptional. Update the nearest guide or specification when changing a documented structure or interface.

Development

Use the root uv project; make install-dev installs the library and development tools into one environment. Run commands with uv run or activate .venv.

Read the full file on GitHub · 35 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 35 lines · 504 tokens per session scan A bdce27798360

Subscribe to this mod's changes

param-decomp CLAUDE.md is an instructions file published in the GitHub repository goodfire-ai/param-decomp (140 stars, last pushed 10d ago), licensed MIT. It adds 504 tokens to every session, about $0.0025 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.