Corpus per-episode stats (the easy ones)

Per-episode descriptive stats across the 12 BBC John Finnemore Double Acts corpus episodes. Computed via analytics.cast_lead + analytics.act_breakdown from the canonical Episode JSONs at data/input_corpus_gold_standard/input_extraction/episodes/*.json.

Per-episode table

ep	title	dur (min)	lines	words	w/l	chars	minor	lead	lead %	2nd %	lead/2nd	scenes	beats	sfx
01	A Flock of Tigers	26.9	379	4990	13.2	2	2	Dolores	50	48	1.03	3	9	6
02	Wysinnwyg	29.2	304	5214	17.1	3	1	Kerry	49	48	1.01	7	10	12
03	Red-Handed	45.5	409	5452	13.3	3	1	Joel	51	47	1.07	4	15	10
04	The Goliath Window	27.2	225	3643	16.2	2	1	Luke	50	49	1.01	4	8	11
05	English for Pony-Lovers	28.7	448	4867	10.9	2	3	Elke	51	48	1.07	3	11	4
06	Hot Desk	28.1	184	4351	23.6	4	0	Mike	46	45	1.02	23	10	53
07	The Queen's Speech †	0.5	314	4472	14.2	2	1	Mabel Nesbit	51	47	1.09	6	12	20
08	Mercy Dash	28.0	507	4983	9.8	3	1	Sue	58	40	1.43	7	10	19
09	The Rebel Alliance	28.1	306	4790	15.7	3	4	Eileen	49	45	1.09	6	16	25
10	Penguin Diplomacy	28.0	378	4341	11.5	3	1	George Bunning	48	48	1.00	7	11	23
11	Here's What We Do	46.7	267	3439	12.9	4	3	Pidge	49	44	1.10	10	16	25
12	The Wroxton Box	28.1	364	4924	13.5	2	1	Alec	50	50	1.00	3	10	43

† Ep 07 duration is a metadata bug in the canonical JSON (28s recorded instead of ~28 min). The line/word counts are real.

Aggregates

metric	min	median	mean	max
duration (min)	0.5†	28.1	28.7	46.7
total lines	184	339	340	507
total words	3439	4829	4622	5452
words / line	9.8	13.4	14.3	23.6
cast size	2	3	2.8	4
beats	8	10.5	11.5	16
sfx events	4	19.5	20.9	53
lead line share	0.46	0.50	0.50	0.58
lead/2nd ratio	1.00	1.04	1.10	1.43

Reading

Lead/second is essentially balanced. The corpus mean lead/2nd ratio is 1.10 — both leads talk roughly equally. Mercy Dash (1.43) and Hot Desk (1.02) are the extremes; even the imbalanced one is only 58/40, not a one-character monologue. Generated episodes should target ~50/50.

Words per line lands at ~13. Median 13.4. Range 10–17 covers 10 of 12 episodes; only Hot Desk (23.6) is an outlier (it has fewer, longer monologue-y lines). Generated content with avg w/l outside 10–17 is suspicious.

Cast size 2–4. Median 3. Anything beyond 4 is non-corpus. The new rule_cast_size_2_to_4 enforces this.

Beats 8–16, median 10.5. Generated episodes typically hit the ~10 mark; the rule exists.

SFX events vary wildly (4–53). Hot Desk + Wroxton Box both > 40 SFX; Pony-Lovers has 4. The system probably over-enforces "few SFX" — corpus actually allows a lot when the form supports it.

Duration clusters at 28 min for all but episodes 03 and 11 (specials at ~46 min). Generated short-form runs at 8 min are about a third the corpus length and should expect proportionally fewer lines/words/beats.

Comparison: v6d vs corpus medians

metric	corpus median	v6d (8 min)	v6d-as-fraction-of-corpus	proportional?
duration (min)	28.1	8.0	28%	—
total lines	339	78	23%	yes (slightly under)
total words	4829	957	20%	yes
words / line	13.4	12.3	92%	yes (just below corpus)
cast size	3	2	—	within 2–4 band
beats	10.5	9	—	normal

v6d is proportionally short-form — line, word, and beat counts scale roughly with duration; w/l is in band. The form is honoured.