Emotional Grammar — Proof Page

A working "periodic table of feeling": 28 entries (25 named human emotions + 3 novel authored coordinates), each defined once as a coordinate in the same 8-axis space and decomposed into atomic elements with a compose-recipe — so a piece can be composed from the formula, not by luck. This page holds a live falsifiability test of that claim.

Full table & map: the 28-entry explainer

The 8 shared axes

Valenceaversive ↔ appetitive

Arousalstill/calm ↔ activated

Tensionresolved ↔ loaded/unresolved

Brightnessveiled/dark ↔ luminous/open

Weightheavy/sinking ↔ light/rising

Motion vectorbackward ↔ still ↔ forward; arrival granted/denied

Breadthintimate ↔ vast

Texturesmooth ↔ agitated

Example entries (atomic elements)

grief-catharsis — tension · weight · valence — arrival: granted

lowered_third / flattened_sixth installed
the learned "sad" interval grammar of Western ears

veiled_resonance installed
muted strings, few open harmonics — sound "behind glass", read as inward/heavy

tension_that_almost_resolves bio+installed
near-resolutions accumulate unspent load the listener wants discharged

sustained_non_resolution installed
nothing fully lands — what-was suspended against what-is-gone

silence_before_release biological
full stop primes the discharge; the pause is the setup, not empty

release biological
resolution after the silence — the cry; catharsis lives in the letting-go

dread — tension · brightness · arousal — arrival: denied

withheld_resolution_indefinite
tension that never discharges — the exact opposite of catharsis; this is the element the A/B test below tries to force

endless_approach
a threat forever nearing, never landing

silence_as_threat
absence where sound is expected — the pause is the predator

triumph — tension · weight · arousal — arrival: granted, fully

earned_tension_release
long accumulation of stakes resolved on arrival

perfect_authentic_cadence
the most conclusive arrival home — victory complete, no doubt left

minor_to_major_brightening
dark→bright modal shift at the turn: overcoming, coded in harmony

The A/B falsifiability test

The real test of a grammar is generative adequacy: compose from the formula and reliably get the target feeling. The grief-catharsis entry claims its release element is load-bearing — "Load-then-release is mandatory; release without prior load does not move." So: two pieces were rendered through the ElevenLabs Music composition-plan rail with byte-identical plans except the final section. Predictions were written down before either render existed.

Piece A — recipe-faithful (release granted)

Veiled B♭-minor ground → slow climb of near-resolving suspensions → peak density → silence → a single warm resolving cadence that lets go.

Piece B — one element inverted (release withheld)

Identical first four sections. Final section replaced: cadence denied, dissonant suspension held, ends hanging — the grammar predicts this stops being grief and becomes dread.

Pre-registered predictions (written before rendering)

P1: A's final seconds fit a resolved triad markedly better than B's; B carries more energy on suspended/non-triad tones.
P2: A decays to rest; B ends with energy hanging or cut.
P3: Both share the same load shape (build → trough → final block) — the pair differs only where intended.
P4: A reads as cathartic sorrow; B reads as dread/suspension.

Measured outcome (honest scorecard)

Pred.	Result	Measurement
P1	MISS	Final-6s best-triad fit: A = 0.699, B = 0.714 — B's ending fits a tonic triad slightly better. ElevenLabs resolved B anyway (both pieces end on bare tonic fifths) despite explicit "cadence denied / no consonant final chord" instructions. The model's cadence bias overrides the inversion.
P2	MISS	Both decay fully to silence (last-3s ≈ 0 for both). Residual difference not in the registered metric: A's dying tail is 13.1 s, B's 7.7 s, and B's music ends 7.4 s earlier — B stops faster, A exhales longer.
P3	PASS (weak)	Both show the identical mid trough at 34.5% of peak before the final block (<40% threshold) — the pair differs only where intended. But no true silence rendered: that atomic element transmitted only as a relative dip.
P4	NOT ASSESSABLE	The audio-perception model (Qwen2-Audio) lives on a GPU box that was owned by another render today; no blind listen was obtained. No claim made beyond the deterministic measurements.

Verdict

Reported as a miss for generative adequacy through this rail: 1 of 3 measurable predictions held, and only weakly. The grammar's claim ("release is load-bearing") was not falsified — it was not expressible: the ElevenLabs composition-plan rail refuses to withhold a cadence, so all arrival-denied entries (dread, longing, ache) currently cannot be A/B-proven through it. Two durable findings: (1) prompt-level negatives do not control endings in this model — withheld-resolution states need the deterministic local-synthesis rail, where the ending is authored per-sample; (2) planned section durations are not honored literally, so all future measurements must locate structure empirically from the audio, never from plan offsets. An honest miss with the mechanism identified.

Receipts

Renders: ElevenLabs Music music_v1, song_ids 3JbW0FA6PjovseFVrJcm (A), gm9bosKcpa8dExfBkNMI (B), 2026-07-06. Plans, pre-registered predictions (PREDICTION.md), analysis JSON, and outcome doc live in art/emotional-grammar/renders/grief-catharsis-ab/. Analysis: librosa RMS envelopes + chroma-triad correlation on the audible content of each file.

Follow-up: the same test, on a rail that can actually withhold a cadence

The ElevenLabs run above came back a rail-limitation MISS: the model resolved the withheld piece anyway, and never rendered true silence. So the same falsifiable claim was re-run on a fully deterministic rail — every note hand-authored in MIDI (C minor, 66 BPM, 14 bars, same veiled → climb → peak-density → silence → arrival structure), rendered through sfizz + VSCO-2-CE sampled orchestra on Foundry (CPU only — the GPU stayed at 100% util on another job throughout, untouched). Sections 1-4 are verified event-identical between the two variants (same RNG seed, same call sequence, checked before rendering); only the final ~5.5s arrival differs.

Variant A — resolves (release granted)

Same veiled-minor → climb → peak-density → silence build. Arrival: Picardy tonic (C major), melody descends D5→C5, authored decrescendo to niente.

Variant B — withholds (release denied)

Identical through the silence. Arrival: V7sus4 held (no third, dominant bass — never resolves to tonic), melody rises to and parks on F5 (scale degree 4), no decrescendo, abrupt cutoff.

Pre-registered predictions (full text: PREDICTION-release-inversion-local.md)

P1: A's final-3s chroma centers on tonic (C); B's doesn't.
P2: Isolated lead line: A lands on C5, B lands on F5, never returning to C5.
P3: A's ending decrescendos to near-rest; B's holds/hangs (open — sample-engine physics could override authored dynamics).
P4: Both variants render a true near-silence window during the authored rest — the capability EL lacked.
P5: A's ending brightens (spectral centroid rises) relative to B's, tracking the introduced major third.

Measured outcome (honest scorecard)

Pred.	Result	Measurement
P1	PASS	Final-3s tonic (C) pitch-class mass: A = 36.7% of chroma energy, B = 3.6%. A best-fits C major (0.799); B best-fits the dominant/sus4 family (G7 0.620, G-sus4 0.618) over tonic (0.614-0.615) — narrow among themselves, but cleanly tonic-poor vs. A.
P2	PASS (exact)	Isolated-violin pitch-track, final 2s: A = 523.25 Hz = C5 to 5 significant figures; B = 698.46 Hz = F5, exactly. No drift from the authored score.
P3	directional PASS, threshold narrowly missed	A's final-3s/peak RMS ratio = 0.202 (true exhale). B's = 0.523 — 2.6x A's, right direction, but under the pre-registered >0.55 bar.
P4	naive window MISS	Beat-math window caught the hall-reverb tail from the preceding forte peak (ratio 0.59) — repeats the prior test's own warning to measure structure empirically, not by plan offsets.
P4 (re-measured empirically)	PASS	True trough found at ~44.8s: RMS = 0.12% of global peak, contiguous near-silence window ~1.47-1.49s, essentially identical in both variants. This is the capability EL lacked, confirmed.
P5	MISS	S3→final centroid: A 471→2446 Hz, B 471→2424 Hz — both rise almost identically. Reverb's diffuse high-frequency tail swamps the harmonic-content signal; not a useful brightness proxy here.

Verdict

The deterministic rail does what the EL rail structurally could not: it withholds a cadence on command (P1, P2) and renders true digital silence on command (P4, once measured at the right place). One mixing-chain artifact was found and named (reverb tail blurs an idealized section boundary by a few hundred ms) rather than an authoring-rail limitation, and one proxy (spectral centroid as brightness) turned out not to discriminate under heavy reverb. Full writeup: art/emotional-grammar/renders/release-inversion-local/OUTCOME.md.