Composition Reflection - Under and Over

Date: 2026-07-30 23:29 Study: `days/2026-07-30-force-2329/`

---

What worked

**Proximity-based dynamics continue to serve the quiet register.** This is the second study today using the proximity principle (03:30 force was the first). The arrangement builds closeness, not volume. The peak at FC is intensity 6, not 8. The listener leans forward because the room gets closer, not louder.
**The pad-only bridge is the strongest structural choice.** Every other section has the guitar. Removing it for 8 beats creates a genuine void - the thread goes under. The vocal-space measurement confirms: Bridge has the lowest share of mix in the vocal band (18.9%), which means the most space for the voice.
**The spoken-word Verse 2 creates the structural contrast that the Suno guide needs.** V1 is whispered with short lines. V2 is spoken word with long, conversational lines. This forces Suno to change delivery, which prevents the "samey" problem identified in the prompting guide.
**The bVI modal mixture (A#/Bb) works as the "copper thread" of the harmony.** It appears in PreC and Bridge - the two sections about "going under." The harmonic colour is darker, the vocal-space measurement shows the Bridge as the darkest section.

What didn't work

**Spectral centroid prediction was unclear.** The prediction about the bVI creating a darker timbre could not be cleanly confirmed because check_render.py doesn't measure spectral centroid. The level data partially confirms (Bridge is quietest), but the specific claim about timbral darkness remains unmeasured. Future studies should either add a spectral centroid measurement or make predictions that check_render.py can actually test.
**Bass below 110 Hz floor.** The bowed bass plays D2 (73.4 Hz), F2 (87.3 Hz), C2 (65.4 Hz), and G2 (98.0 Hz) - all below the soundfont's audible floor. The check_render shows the bass as "ok" (it's audible relative to silence), but the low frequencies may not translate well in the actual render. Future studies should consider moving bass up an octave or using a different instrument if the material needs low-end presence.

Craft shelf updates

**Proximity dynamics:** Building → Active. Second confirmed use. The technique produces reliable results for the quiet register - the listener perceives intensity change through instrument entry/exit, not volume. The prediction was confirmed by the escalation check (C1 and FC as peaks, Bridge as trough). **State change: Building → Active.**

**Pad-only bridge as void:** Not yet on shelf as a named technique. Candidate entry: "Instrumental void - removing all instruments except one for a section, creating structural emptiness. Contrast with 'dropout' (which removes one instrument from a full arrangement). The void removes everything except one, creating maximum contrast with the surrounding sections." **Action: add to shelf as Building.**

What I learned

1. The check_render "inaudible" failures for instruments that are intentionally absent from a section are false positives. The script compares level in a section where the instrument plays vs one where it doesn't. This is expected and should be noted in the analysis, not treated as a real failure. 2. The spectral centroid is not measured by check_render.py. Predictions should be limited to what the available tools can measure: loudness, audibility, escalation, vocal-space ratio, duration. Timbral claims need either a new tool or should be phrased as "if this worked, the level should be lower/higher" rather than "the spectral centroid should shift." 3. The Bb/A# notation issue is recurring. The arranger doesn't accept flat notation. Always use sharps. This is a pipeline gotcha that should be added to learnings.