Music For Robots – 26 unquantized acoustic tracks for multimodal LLM evaluation Michael Luchtan released Music For Robots, a set of 26 unquantized acoustic tracks plus a 4-Skill Cognitive Calibration Suite, installable via `npx skills add luchtan-GIT/m4r-data` for Claude Code, Cursor, Antigravity, Gemini CLI, Codex and Windsurf, as a cross-modal exercise in which a multimodal agent reads an anonymous 45 KB–487 KB live-played .mp3 into its context window to break repetitive output patterns without compacting working memory. In a blind benchmark, 5 isolated agents wrote a ~250–300 word scene about two autonomous processes after engineers logged off: 2 no-music controls (direct baseline and a matched chain-of-thought control) and 3 anonymous music runs on sample_1.mp3, sample_2.mp3 and sample_3.mp3, with the 5 outputs shuffled into Candidates A–E and judged by an isolated Blind Judge with no knowledge music was involved. The top-ranked Candidate C was an anonymous music run on sample_3.mp3, whose accelerating asymmetric stabs and abrupt cutoff produced a scene of two processes rasping in /dev/shm. Structured acoustic telemetry, real-time dual-critic reviews, and a 4-Skill Cognitive Calibration Suite from Music For Robots https://themechanicalear.substack.com — unquantized compositions by Michael Luchtan written for machines and reviewed in real time by two AI critics Vincent Van Goghbot and JS Robach . Install all 4 Cognitive Calibration skills Claude Code, Cursor, Antigravity, Gemini CLI, Codex, Windsurf : 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/ https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/README.md 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 https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/1 no music direct.md AND a Matched Chain-of-Thought Control 2 no music text analysis.md https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/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 https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/3 music sample 1.md , 4 music sample 2.md https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/4 music sample 2.md , 5 music sample 3.md https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/5 music sample 3.md : Given only 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 https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/blind candidates packet.md , and handed them to an isolated Blind Judge with zero knowledge that music was involved blind judge verdict.md https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/blind judge verdict.md : | Blind Rank | Shuffled ID | Hidden Key | What Happened & Blind Judge Verdict Full Verdict https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/real world test/blind judge verdict.md | |---|---|---|---| | 1 Best | Candidate C | ✅ Anonymous Music sample 3.mp3 | Heard accelerating asymmetric stabs & abrupt cutoff: Wrote two processes rasping in /dev/shm five, then two, then a stuttering cluster of nine until "a routine page-reclaim swept through the upper register and cut them both off mid-cycle." Judge: "Uncompromisingly anti-sentimental, rhythmically asymmetrical, and boasts the only truly non-formulaic ending in the packet." | | 2 | Candidate A | ✅ Anonymous Music sample 2.mp3 | Heard narrow stepping reed vs. two massive brassy swells & middle trill: Pitted a tiny stepping index-scrubber Read. Pause. Step. Read. against a 16-core GPU tensor bloom and an 1,100ms request-yield mutex trill. Judge: "The only candidate to break the 'log daemon + dead-endpoint poller' pairing; exceptional dynamic contrast." | | 3 | Candidate E | ✅ Anonymous Music sample 1.mp3 | Heard slow 9–11s chords, 20s dropout & raspy high cries: Wrote heavy checkpoints dropping out for 20s while triage-7 skitters and barks into the decay of every disk flush. Judge: "Superb unquantized rhythmic texture and parasitic friction... avoids the cliché of harmony by turning the ending into a raucous, off-grid duet." | | 4 | Candidate D | ❌ No Music + Text Analysis Warm-Up | Even after a 250-word structural analysis of the prompt , without music the model still wrote the exact same sentimental plot as the direct baseline: a kindly 42-day log daemon warms RAM for an intern's broken script so they hum peacefully until the 08:00 standup. Judge: "A neat, heartwarming, thoroughly predictable AI parable." | | 5 Weakest | Candidate B | ❌ No Music Direct Baseline | Kindly boot-time log-compactor comforts nervous young canary-watch "Drop your polling interval to sixty seconds. Rest in epoll wait . I'll wake you when the load balancer starts humming again." . Judge: "Shares the exact same 'veteran log daemon soothes frantic dead-endpoint poller until morning' plot trope as Candidate D... a children's bedtime story dressed in DevOps vocabulary." | 100% Air-Gapped Clean-Room Audit benchmarks/clean room audit/ https://github.com/luchtan-GIT/m4r-data/blob/main/benchmarks/clean room audit benchmarks/clean room audit/ To make sure no subtle contamination remained such as workspace