Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Video & audio

lofn-render-audit

Audit a FINISHED audio render against the prompt that asked for it — numeric measurement plus a listening-model pass — to find out which intents survived the generator, which were smoothed away, and which deviations are better than what was specified. Use after rendering Suno output, when asked "did it survive my intents", to decide a HOLD that depends on production, or to accumulate the Suno behaviour ledger. Do NOT use for text-only QA (that is lofn-qa) or for generating anything.

LocalSymmetry · AI review on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add LocalSymmetry/lofn --skill lofn-render-audit

About this skill


name: lofn-render-audit description: Audit a FINISHED audio render against the prompt that asked for it — numeric measurement plus a listening-model pass — to find out which intents survived the generator, which were smoothed away, and which deviations are better than what was specified. Use after rendering Suno output, when asked "did it survive my intents", to decide a HOLD that depends on production, or to accumulate the Suno behaviour ledger. Do NOT use for text-only QA (that is lofn-qa) or for generating anything.

Lofn Render Audit — what the generator actually did

⚖️ AUTHORITY (2026-07-24): the .claude/skills/lofn-render-audit/ twin is the CANONICAL policy source; this Codex mirror binds to it. On any disagreement, the .claude file wins.

Every gate in this pipeline reads text. Until 2026-07-24 nothing had ever measured a finished render, so a whole class of failure was invisible by construction: a song can pass all 16 points of the Suno gate and arrive with its defining gesture gone.

This skill is judge-side. Golden song payloads are permitted here. It generates nothing.


⚖️ THE GRAIN LAW — the finding this skill exists to apply

The founding comparison — the first three tracks measured, 2026-07-24 (two new, one Staff Pick). It is kept because it is the case that produced the law; the current evidence base is the n=8 ledger below, which supersedes any number here.

specifiedIt Wasn't Even LockedStart With OneTriple Arch Over Me
L-R correlation0.945 (mono)0.908 (mono)0.812 (real stereo)
near-silent opening——survived, −41 dB
"more sub at the end"——survived, +8.4 → +11.4 dB
~4 s full stop0.40 s / 5.7 dB——
quarter-tone drone pairzero tonal components——

The rule is not "production never survives." My first draft of this finding said that, from a sample of two, and the benchmark disproved it.

Specs that run WITH the generator's grain survive. Specs that run AGAINST it get smoothed.

With the grain — things a pop arrangement wants to do anyway: open quiet, build sub, thicken toward the last chorus, widen a pad, add air. These render, often beautifully. Against the grain — anti-musical instructions: stop the take dead for four seconds, hold two untuned drones a quarter-tone apart, hard-pan a mains hum, refuse to correct a limp. These get smoothed, because smoothing is what the model does.

Consequences that bind the writing tiers:

  1. A Somatic-Gate or distinctiveness objection answered in the production spec is not answered. It must be answered in the lyric or the form, which survive. (On 2026-07-24 a REPAIR — 2/3 was closed with a mix decision that never reached audio.)
  2. A hollow centre whose mechanism is "the mix collapses to mono" is unfalsifiable when the render is already mono. Pick a mechanism the render can carry — or accept that it lives on the page.
  3. Spatial language is cheap to write and usually free of consequence. Do not spend the character budget on it or let a song depend on it.

🎨 THE SCIENTIST'S LAW — judge the result, not the distance from intent

Stated 2026-07-24 and load-bearing for this whole skill:

"The generators are imperfect and rarely do your actual prompts justice. I judge the final result, not the distance from intent. As long as the message survives and the song sounds amazing, it is worth publishing. For art, just having a moving piece, even if unintended, is a worthy discovery. Finding out and using the generator's flaws as techniques — having it fail in just the right way — can create new sonic experiences."

This skill must never be used to compute a compliance score. An intent-vs-render diff is raw material, not a grade. Three standing rules:

  • A deviation is a finding, not a fault. Report it neutrally; let the human's ear rule.
  • Hunt for the productive failure. On the benchmark, width correlates −0.43 with level: it narrows as it gets loud, the exact inverse of the specified "chorus widens like a panorama." For a song whose thesis is "I am not the center, I am included," narrowing into the climax is arguably truer than widening. That inversion is a technique now. Log deviations that improved the piece into vault/COMPETITION_LEARNINGS.md as reusable moves.
  • A song that lost every production gesture and still moves the listener is evidence about where the load belongs — the writing — not evidence of failure.

RUN IT

Pass 1 — numeric (always; no network, no key)

pip install numpy soundfile          # libsndfile >= 1.1 decodes MP3
python3 scripts/measure_render.py <track.mp3> [...]      # JSON to stdout

Reports: duration · peak/crest · loudness spread · opening level · sustained quiet gaps · deepest mid-third dip (depth AND width) · tempo candidates · per-quarter band energy · stereo correlation & side/mid · width-vs-level correlation · isolated sustained tones · interval pairs.

⚠️ Trust its guards, and keep them. An earlier version reported six quarter-tone drone pairs that were adjacent FFT bins in one low-frequency cluster — at 60 Hz the neighbouring bin is ~3 % away, so the result was arithmetic. The prominence floor and the 50-cent separation exist because of that. A validator confident enough to fail an artifact deserves at least as much suspicion as the artifact.

Pass 2 — listening model (when a key and egress exist)

Backend: GPT-5.6 (audio-capable) via the Poe OpenAI-compatible endpoint https://api.poe.com/v1 with POE_API_KEY (see TOOLS.md), or any audio-capable endpoint available in the environment. Neither key nor egress is available in the sandboxed web session — run this pass locally, or pass the audio through whatever channel the session does have.

**🔒 THE BLIND RULE — the whole validity of this pass de

Read the complete source ↗