code/CAPTION_GATE.md

caption_general — the emotion gate and the GEND/BKGN polarity

Regenerated corpus-wide on 2026-08-23 over 165,516,420 rows / 48,556 shards. The previous string is preserved verbatim in caption_general_v1 on every row, so the change is fully reversible and auditable.

Emotion clause — percentile gate, not an absolute threshold

The old clause named an emotion when its raw Empathic-Insight score cleared an absolute emo_thr = 1.0. The 40 heads are not on a common scale, so that gate reported the scale of the head rather than the emotion of the clip:

head	median	zeros	named before
emo_Interest	2.082	0.0 %	90.6 % of all rows
emo_Concentration	1.543	6.1 %	71.0 %
emo_Infatuation	−0.017	87.7 %	0.3 %
emo_Bitterness	0.012	5.5 %	0.1 %

The rule now: an emotion is named when it falls in the top 10 % for that emotion (U ≥ 0.90), at most 3 named, ranked by percentile. A clip that clears nothing says "no dominant emotion" rather than being forced to pick.

The normaliser is capnorm.npz: a pooled, tie-aware mid-rank ECDF over 132,833,726 utterances (traj2/stats/globalhist.npz, 130,785,282 rows across 8 datasets and every language, plus vprof_vc fitted into the same 4096 bins). It is not re-fitted here — an earlier sampled-knot ECDF gave emo_Awe only 120 distinct knots and made it the "peak emotion" of 23 % of clips.

Pooled, deliberately not per-dataset. Under the pooled norm the datasets genuinely differ (mls 0.615, evasnippets 0.608, eurospeech 0.547, emolia 0.441). Per-dataset normalisation would force every dataset to 0.500 by construction, erase that signal, and put captions on a different scale from the trajectory miner, which selects chains with the same statistic.

Measured effect corpus-wide: Interest 90.6 % → 5.3 %, all 40 emotions now occur, mention count 428.3 M → 341.1 M, and 17.88 % of rows are neutral. The neutral rate is strongly dataset-dependent and that is the intended behaviour of a pooled norm — see the per-dataset table in capfix/out/report.json.

GEND and BKGN polarity

Both ordinal ladders ran backwards relative to the data and were corrected in caption2.py on 2026-08-22 (md5 ec55b223…); this pass applied the correction to the corpus rows.

• High vn_GEND is masculine: +0.842 with chest resonance, −0.484 with head resonance, −0.352 with brightness. Independently, a sibling classification of all 500 voice profiles from the numeric vn_GEND against each profile's own design-spec card_gender agreed on 89.2 % of 379 decided voices — an inverted ladder would have scored ~11 %.
• High vn_BKGN is cleaner: +0.786 with recording quality and +0.29 with qual_background_quality, an independent model head. The old wording produced the self-contradictory "good recording, very noisy background"; it now reads "good recording, no background noise".

The corrected tag is recomputed from vn_GEND_bucket / vn_BKGN_bucket, not by flipping the string, so the operation is idempotent and safe to re-run. This mattered: the -reann tree (vprof_base + vprof_repaired, 28.2 M rows) had already been regenerated with the corrected ladders and measured 100 % new polarity, while all nine live-tree datasets measured 100 % old polarity. A blind string flip would have re-inverted those 28.2 M rows.

What did NOT change

Only three clauses may move: the emotion clause, the GEND token inside clause 0, and the BKGN token in the recording clause. Delivery, timbre, speech, affect, style, recording quality, the explicit-content flag, burst handling, text_with_bursts, genuineness, blend, duration and language are carried across byte-identical. This is asserted per shard during the rewrite and re-checked afterwards: other_clause_moved = 0 over all 165,516,420 rows.

One wording change does fall inside the emotion clause: emo_Jealousy_and_Envy renders as "jealousy and envy" where the old renderer wrote "jealousy & envy".