A–G taxonomy

The NEXORA annotation guide describes 7 observation categories (letters A to G), each with coded labels (e.g. A1). Each 15 s window carries one label per category; each label has a base score /10 (see scoring.py::BASE_SCORE).

The 7 categories

Cat.	Name	Labels	Examples
A	Attention Orientation	A1–A5	on-task gaze (`A1`), off-task (`A2`), peer gaze (`A3`), prolonged stare (`A4`), head down (`A5`)
B	Task Activity	B1–B5	writing (`B1`), reading (`B2`), hand raise (`B3`), idle (`B4`), off-task interaction (`B5`)
C	Posture	C1–C5	upright attentive (`C1`), leaning forward (`C2`), slouched (`C3`), head on desk (`C4`), arms crossed (`C5`)
D	Physical Activity	D1–D5	yawning (`D1`), fidgeting (`D2`), nodding (`D3`), stretching (`D4`), no movement (`D5`)
E	Facial / Affective	E1–E5	neutral (`E1`), positive affect (`E2`), confusion (`E3`), frustration (`E4`), face obscured (`E5`)
F	Peer Interaction	F1–F4	none (`F1`), task talk (`F2`), off-task talk (`F3`), physical contact (`F4`)
G	Teacher-Cue Response	G1–G4	reorients to teacher (`G1`), no response (`G2`), delayed (`G3`), no cue (`G4`)

Scoring rules

Window = mean of the 7 category scores (unless overridden):

window score = (A + B + C + D + E + F + G) / 7

Contextual rules (a label's score may depend on other categories):

Rule	Condition	Result
A3 resolved by F	`A3` + `F2` (task talk)	A = 10
	`A3` + `F3` (off-task talk)	A = 3
A5 resolved by B	`A5` + `B1/B2` (writing/reading)	A = 9
	`A5` + `B4/B5` (idle / off-task)	A = 0
	`A5` unresolved	A = 5
B4 adjusted by A	`B4` + `A2/A4` (off-task gaze)	B = 2
	otherwise	B = 5
G2 override	`G2` (no response to a cue)	window = 0 (always)

Default labels (rule R2)

When a category was not annotated on a window:

A → A2 (unverifiable gaze)   B → B4 (idle)   C → C1 (standard posture)
D → D5 (no movement)         E → E1 (neutral) F → F1 (no interaction)   G → G4 (no cue)

Posture C is sampled every 60 s then propagated (forward_fill) over the intermediate windows.

Base scores (extract `/10`)

Label	Score	Label	Score	Label	Score
A1	10	B1/B2/B3	10	C2	10
A2	2	B4	5	C1	8
A4	1	B5	0	C3	3
A3 (solo)	3	D5	7	C4	0
		D3	10	C5	5
E2	10	D2	3	F1	8
E1/E5	6	D1	2	F2	10
E3	5	D4	4	F3	2
E4	2			F4	4
G1	10	G2	0	G3	5 · G4

Input tolerance

The pipeline accepts free text (human CSV) and normalizes it: category variants ("physical activity signals" → D), punctuation/case ignored, aliases ("on task" → A1, "writing" → B1…). What cannot be normalized goes to silver/rejects.