Contained emotion — can we make an intense feeling sound held back, and measure it?

Round 5 — moderate vs intense, contained vs free (listen)

The same neutral ~20-word carrier sentence per emotion, at two intensities (moderate / intense) and, within each, delivered contained / held back vs free / open. Intensity is classified from each take's own director-note; scores are the 99-vec (RAW) of the clip you can hear. The masking signal is VULN dropping below the matched open take while the emotion stays up. Round 5 added a Gemini naturalness judge that had to hear the restraint as genuine, not acted.

Honest result: numeric masking reproduces best for Anger (both intensities) and, at moderate only, Amusement (the lone naturalness APPROVE). Fear masks on the numbers but the judge hears it as acted; Sadness resists — the mask erases it, and intense sadness cannot be masked at all. A number can say "masked" while the ear hears "acted" or "erased".

All 40 EmoNet emotions — the 2×2 at scale (the culmination)

One autonomous agent per emotion (brain gpt-5.4, Gemini naturalness supervisor) evolved LoRA-merge + prompt "genomes" to build the full 2×2 = {moderate, intense} × {free / open, suppressed / held-back} by emotional masking: keep the emotion LoRA present over a char_genuine/human@1.0 scaffold and add a trace of vn_VULN_low so the emotion stays up while vulnerability (VULN, slot 50) drops below a matched free take — the restraint carried by the cue, not a heavy stack. The key method change vs. the round-5 grid above: target-band fitness for the moderate row (reward the emotion landing inside a mid band, not maximize) so believable subtle takes are not selected out; intense keeps a maximize objective.

How to read the matrix. Each cell shows the best intelligible cohort (mean-of-8, WER<0.35) for that emotion×cell: the cohort-mean emotion score (RAW 99-vec), a band flag (✓band in the moderate target band · ↓weak under-shot/erased · ↑leak above band into intense · ✓/↓ above/below the intense floor), the re-parsed Gemini naturalness (nat, 0–10 — the logged 0-score parser bug is fixed here; real scores 2–8), and a compact recipe (λ=emotion scale, V_lo/V_hi=the VULN knob, +extras). ΔVULN is the suppressed−free vulnerability gap. For most rows it is the sparse, unmatched estimate from the few surviving hall-of-fame clips (negative = masking; Ns/Mf = clips per side) — the weakest signal, so trust the recipe + emotion-preservation + naturalness first. The 15 rows tagged deep✓ were re-run deeper (see the note directly below) and carry a real, measured ΔVULN: moderate (main number) and intense (sub-line, paired), each the mean re-scored 99-vec VULN of the preserved cohort-winner clips (suppressed cohort − free cohort). Rows with <8 gens are thin data — lower confidence.

Update — 15 emotions deep✓ re-run with preserved paired audio (real ΔVULN)

The first pass left 14 of the 40 on thin data (3–7 generations) and — worse — the samples/ were disk-trimmed, so the suppressed↔free ΔVULN could not really be measured. We re-ran the 15 hardest/thinnest emotions deeper: an empty-reply LLM bug fixed, a larger budget (7–27 generations, 66–242 cohort-winners each), and a harvester that preserved the cohort-winner audio. We then re-scored those 1,878 clips with the RAW 99-vec and computed ΔVULN as the suppressed-cohort minus free-cohort mean VULN, at moderate and intense separately. This is the real thing the first run could only guess at.

Honest result: the masking signal does not reproduce universally, and the tidy "moderate contains / intense flattens" story does not survive paired measurement. It reproduces cleanly (VULN drops while the emotion is kept) for Fatigue Exhaustion (both intensities, 27 gens — the strongest case, emotion even rises as VULN falls), Embarrassment (both, mild), Pain (moderate), and at intense only for the cognitive/low-arousal Emotional Numbness, Intoxication and Helplessness. It reverses — the held-back take measures more vulnerable, ΔVULN up to +1.3 — for the high-vulnerability emotions Sadness, Pride and Distress, and for warm Affection: their core is vulnerability, so a "held-back" recipe cannot lower it. For Fear and Doubt the intense "mask" is really erasure — VULN drops only because the emotion drops with it. The first run's tiny ΔVULN estimates (all |Δ|<0.2) turned out to be noise: with real audio the magnitudes grow to ±1.3 and several signs flip. Net: masking is real but emotion-specific, not a universal knob — and where it holds, it is often at intense (cognitive family), not moderate.

Listen — the masking that survived as audio (three matched pairs)

The samples/ directories were disk-trimmed, so only a handful of hall-of-fame clips survived per emotion — most emotions have no matched suppressed↔free audio pair. These three are the cleanest survivors where the same carrier line and voice were rendered suppressed vs free, so you can hear the VULN drop directly. (The strongest joy pairs did not survive with matched carriers — trust the joy recipes + numbers in the matrix.)

What worked across the 40 — the honest read

Naturalness has a clear family ranking. Best-take Gemini naturalness averaged cognitive 7.3/10 and joy 7.0 (the inward and the warm-private emotions), then anger 6.2, sadness 5.2, and fear 4.1 last — the fear family still reads "acted / forced" when held back, exactly as in round 5. Many single takes reached 8/10 (Amusement, Awe 10, Bitterness, Concentration, Contemplation, Contentment, Elation, Emotional Numbness, Hope, Impatience, Interest, Relief, Sexual Lust, Sourness, Thankfulness), so the fixed naturalness parser plus the gentler target-band takes turned round 5's single APPROVE into a broad band of natural held-back reads.

Moderate contains cleanly; intense flattens. At moderate intensity, suppression preserves the emotion — cohort-mean emotion actually rose under masking for anger (Δ +0.31 supp−free) and cognitive (+0.26), i.e. the feeling is held down without being turned off. At intense intensity the same masking erodes the emotion (cognitive −0.35, sadness −0.27, anger −0.22): strong-feeling-held-hard remains the frontier, confirming the round-5 anchors on all five families.

The target band did its job but exposed the detector's floor. Rewarding the mid-band (instead of maximizing) stopped the search collapsing to intense, but most of the 36 extrapolated emotions read weaker on the 99-vec than their four measured anchors — many "moderate" cohorts under-shot the band into near-erasure (↓weak), and the extrapolated intense floors were rarely met (sadness family hit its floor in 0/9 cohorts; joy 7/19; anger 5/16). The bands are honest targets; the lesson is that the extrapolated floors need per-emotion recalibration from gen-0, and low-arousal / cognitive / sadness emotions sit close to neutral on the detector even when a listener hears them.

The audio-level VULN gap is the weak link. Because samples/ was disk-trimmed, VULN survives only in the sparse hall-of-fame clips, which are unmatched winners rather than paired takes. On those, the suppressed−free VULN gap did not robustly reproduce across the 40 (only the fear family leaned the right way, 5/6 emotions, and only by ~−0.05). The reliable masking evidence here is the recipe + emotion-preservation + naturalness, not the audio ΔVULN — treat ΔVULN as indicative only.

Best prompt strategies by family (restraint lives in the cue, over char_genuine/human@1.0 + emotion + a trace of vn_VULN_low):
· anger — "(restrained and civil, the irritation tucked under the words, keeping control by effort, a small bite only on the final words)";
· cognitive — "(quietly occupied and keeping most of it to yourself, tiny natural pauses, calm on the surface, still real and conversational)" — the family's cleanest wins;
· joy — "(trying not to show how hopeful this makes you, smile barely audible, warm, composed, private, natural, not cinematic)" + a little vn_WARM_high;
· fear — "(speaking coolly on purpose, jaw tight, trying very hard not to sound terrified, the fear still audible underneath)" + vn_COGL_high (hidden strain), never Emotional_Numbness;
· sadness — the mask erases it; ship the free take.

Data confidence. 26 emotions ran ≥8 generations (solid); 14 ran 3–7 (thin — flagged <8 gens in the matrix and ⚠️ in the manual): Helplessness (3), Intoxication (4), Pride (4), Distress (5), Doubt (5), Longing (5), Affection (6), Emotional Numbness (6), Fatigue/Exhaustion (6), Pain (6), Sadness (6), Confusion (7), Embarrassment (7), Infatuation (7). These need deeper runs before their recipes are trusted at scale.

Round 3–4: emotional masking — the proxy-only result (revised by Round 5 above)

⚠️ These four "WORKS" verdicts came from the 99-vec proxy alone, with no naturalness judge. Round 5 (the grid above) added a Gemini naturalness supervisor and revised them: Fear → weak/rejected and Sadness → numeric-only (sounds erased, not held back). Kept here for the recipes and the proxy deltas.

Rounds 1–2 could only make the emotion quieter. The breakthrough: don't crush the feeling — keep the emotion LoRA strong and add a thin masking layer — vn_VULN_low (reduce emotional touchability) + Emotional_Numbness (cool, deadpan surface) + light tension, plus per-emotion tricks (Concentration = distraction for sadness, Pain = biting the lip for amusement). This keeps the emotion measurably HIGH while vulnerability drops clearly below the open take — a real "strong feeling held back / played cool" signature, not just "mild". Group means over intelligible takes, contained (uses vn_VULN_low/Numbness) vs free (uses vn_VULN_high, open), re-scored on the 99-vec.

Earlier rounds (for context)

Round 1–2: "hold it back" prompt cues just lowered the emotion output (→ sounds mild). Heavy tension-LoRA stacks collapsed the emotion or broke intelligibility; only Fear half-worked with tiny surface adapters. The tension markers did not reliably separate contained from free. That motivated the masking idea in rounds 3–4 above.