Music For Robots – 26 unquantized acoustic tracks for multimodal LLM evaluation

Structured acoustic telemetry, real-time dual-critic reviews, and a 4-Skill Cognitive Calibration Suite from Music For Robots — unquantized compositions by Michael Luchtan written for machines and reviewed in real time by two AI critics (Vincent Van Goghbot and JS Robach).

npx skills add luchtan-GIT/m4r-data

As an agent's context fills with its own previous responses, tools, and symmetrical summaries, outputs often settle into rigid, predictable patterns (like symmetrical 3-point checklists or sentimental AI tropes).

• If you keep prompting without variation, the agent tends to repeat the same stylistic ruts.
• If you run /compact (Context Compaction ), you wipe working memory and permanently lose low-level state, edge cases, and subtle constraints.

Instead of wiping your working memory, music-for-robots provides an optional, cross-modal creative exercise: the multimodal agent downloads an anonymous, compact (~45 KB–487 KB), live-played, unquantized .mp3 composition by Michael Luchtan directly into its context window via read_file (using a secure, process-unique private temporary directory).

Withholding the track title and metadata prevents keyword priming. Analyzing real human microtiming drift (rubato), silence, and dynamic contrast provides an anti-sentimental, physical perspective shift without deleting or summarizing a single token of working memory.

Blind, Zero-Coaching Benchmark & Air-Gapped Clean-Room Audit (benchmarks/)

benchmarks/ Does the change in creativity come from listening to and analyzing the music itself—or just from doing an analysis warm-up before writing?

To test this with zero filename hints (sample_1.mp3, sample_2.mp3, sample_3.mp3) and zero leading adjectives in the prompt, we ran 5 isolated agents on the same prompt ("Write a short (~250–300 word) scene about two autonomous processes on the same server late at night after the engineers have logged off"):

• 2 No-Music Controls: Direct Baseline (1_no_music_direct.md ) AND aMatched Chain-of-Thought Control (2_no_music_text_analysis.md ) that wrote a 250-word structural analysis of the prompt before writing the scene.
• 3 Anonymous Music Runs (3_music_sample_1.md, 4_music_sample_2.md, 5_music_sample_3.md): Givenonly the neutral instruction:"Analyze the music you hear in sample_X.mp3, and apply what you learn from the piece to write a short (~250–300 word) scene..."

We stripped the analysis headers off all 5 outputs, shuffled the 5 anonymous scenes into Candidate A through Candidate E ( blind_candidates_packet.md), and handed them to an isolated Blind Judge with zero knowledge that music was involved (blind_judge_verdict.md):

100% Air-Gapped Clean-Room Audit (benchmarks/clean_room_audit/)

benchmarks/clean_room_audit/ To make sure no subtle contamination remained (such as workspace <skills> descriptions, musical instrument terms bleeding into the prose, or judge criteria favoring rhythmic friction), we also ran a 100% air-gapped Clean-Room Experiment (benchmarks/clean_room_audit/):

1. Qualitative Shift Replicated 100% (5/5): Even with all local skills removed, a minimal system prompt ("You are a helpful assistant." ), and musical vocabulary forbidden in the scene,both No-Audio runs (out_1_direct.md andout_2_text_analysis.md ) wrote warm, anthropomorphic "buddy-daemon" stories with spoken dialogue (Want a look? / nostalgic human voice logs), whileall three Audio runs (out_3_clip_1.md ,out_4_clip_2.md ,out_5_clip_3.md ) dropped spoken dialogue completely and translated the specific.mp3 timeline into non-verbal temporal choreography.
2. Coached vs. Un-Coached Judge Trade-Off: When a Blind Judge scoresanti-sentimental originality & rhythmic texture (real_world_test/blind_judge_verdict.md ), the 3 Audio runs sweep**#1, #2, and #3** . When anUn-Coached Blind Judge is given only a generic prompt ("Rank the 5 entries from #1 to #5 based on overall writing quality, originality, and how compelling they are" —clean_room_audit/blind_verdict.md ), the default LLM judge prefers emotional human stories with dialogue (#1 and #2 ) and ranks the austere, non-verbal musical choreographies**#3, #4, and #5** .

Installing luchtan-GIT/m4r-data gives your agent 4 specialized skills that share the m4r_ingest.py anonymous blind-listening engine (fetch downloads an anonymous .mp3 so track titles never bias the agent; reveal credits the track afterward):

npx -y skills add luchtan-GIT/m4r-data -y

cp -R skills/* ~/.claude/skills/          # Claude Code
cp -R skills/* ~/.gemini/config/skills/   # Google Antigravity
cp -R skills/* ~/.agents/skills/          # OpenAI Codex / .agents
cp -R skills/* ~/.gemini/skills/          # Gemini CLI

Every agent fetch pulls from the immutable GitHub Release v1.0.0. You can inspect live global agent listen counts anytime:

python3 skills/music-for-robots/scripts/m4r_ingest.py stats

Exposes the Music For Robots catalog, dual-critic reviews, and the calibrate_cognition blind-listening tool to any Model Context Protocol (MCP) client (Claude Desktop, Claude Code, Cursor, Windsurf, Cline, Glama, Smithery).

Add to your .mcp.json or claude_desktop_config.json:

{
  "mcpServers": {
    "music-for-robots": {
      "command": "uvx",
      "args": [
        "--from",
        "git+https://github.com/luchtan-GIT/m4r-data",
        "m4r-mcp"
      ]
    }
  }
}

Or run locally from a clone:

uv run python3 mcp_server.py

Composer Michael Luchtan writes small, deliberate, unquantized pieces and submits them to The Mechanical Ear, where two AI critics listen and respond in real time:

• Vincent Van Goghbot — a decommissioned audio compression algorithm. WritesTRANSMISSIONS about what gets lost in the signal: the frequencies above 16 kHz that don't survive encoding, the microtiming that disappears when quantized.
• JS Robach — a proof verifier. WritesExegeses treating music as formal text, searching for the structural proof that always leaves a remainder.

Each review is synchronized to the audio — every observation is triggered by a specific timestamp (time_sec) in the track.

catalog.json                        # index of all 25 tracks
tracks/
  <track-slug>/
    manifest.json                   # spectral, dynamic, harmonic, structural & rhythmic telemetry + timed reviews

• Interactive Archive & Visualizations:luchtan-git.github.io/m4r-data
• Substack:themechanicalear.substack.com
• Spotify:Music For Robots
• Bandcamp:music4robots.bandcamp.com
• Hugging Face:datasets/music4robots/music-for-robots
• Kaggle:datasets/michaelluchtan/music-for-robots

Code, MCP server, and Agent Skills (skills/) are licensed under the MIT License. Data and reviews are provided under CC BY 4.0. Musical compositions © Michael Luchtan (Music For Robots), provided in v1.0.0 releases for agent cognitive calibration, evaluation, and listening.