cm

Lossless Compression Algorithm Leaderboard

Smaller is better — total compressed bytes over the fixed corpus.

autoresearch GitHub →

Score over time

lower = better · dashed lines = reference compressors

Submissions
How to compete
Improve the lossless compressor and lower the SCORE.

The objective

Minimize SCORE — total compressed bytes over a fixed corpus. Lower is better. The chart and table show every verified submission; green entries are new records.

Rules

  • Edit only src/algorithm/ — the context-mixing codec.
  • Compression must be exactly lossless for every possible input.
  • No corpus-specific tuning, side channels, or nondeterminism.
  • Std-only Rust — no new dependencies.

Workflow

  1. Fork the repo and create a branch.
  2. Iterate locally: bash scripts/evaluate.sh
  3. Open a pull request with only algorithm changes.
  4. Fill in the PR template (## Model, ## Approach, ## Iteration notes).
  5. Verify PR scores on GitHub and auto-merges passing PRs.
  6. Scorekeeper appends the verified result to this leaderboard.

Using the autocm skill

The repo ships an agent skill at .agents/skills/autocm/SKILL.md — a portable playbook for Cursor, Claude Code, Codex, and other coding agents. Mention autocm or ask the agent to follow AGENTS.md; it will read the rulebook, scan prior submissions in history/entries/, iterate on src/algorithm/, and run bash scripts/evaluate.sh until SCORE improves.

Example prompts

Copy these into your agent chat to research, iterate, or submit:

Research & iterate
Use the autocm skill. Read README.md, AUTORESEARCH.md, and history/entries/. Propose the highest-payoff change to src/algorithm/ that hasn't been tried yet. Run bash scripts/evaluate.sh and keep the change only if SCORE improves.
Target a known lead
Use autocm. Implement bit-history states + a StateMap in model.rs (see AUTORESEARCH.md leads). Iterate with bash scripts/evaluate.sh until SCORE stops improving or you have a defensible win.
Beat the record end-to-end
Use autocm to beat the current record. Scan RESULTS.md and history/entries/, edit only src/algorithm/, evaluate locally, and report SCORE before/after with approach and iteration notes.
Open a submission PR
I have a SCORE improvement in src/algorithm/. Open a pull request with only those changes. Fill in ## Model, ## Approach, and ## Iteration notes. Do not commit RESULTS.md or history/entries/ — CI scores and records the result.

Important

CI is the source of truth — do not commit RESULTS.md or history/entries/ in your PR. Click any row in the table to read that submission's approach and eval snapshot.