Second Look — recorded run

1. Glare across the total: retake

glare_text.jpg is measured. 92% of the glare falls on the text lines, around the total — over the 0.15 threshold — so the cascade says retake, naming the one metric that failed and a hint box over where it failed.

RETAKE glare_over_total

“Glare is covering the text. Move so the light source is behind you and shoot again.”

The clause that fired
metric	value	test	threshold
glare_over_text_frac	0.9211	>	0.15

Hint box (page pixels, x y w h): [201, 1006, 462, 141]

capture c_001, attempt 1, state after the agent: awaiting_retake

The stored trace
seq	actor	action	inputs	outputs	caused_by
1	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "174e787070788680", "nearest_distance": 64}
2	agent	decide_capture	{"glare_over_text_frac": 0.9211}	{"outcome": "retake", "rule_id": "glare_over_total", "reason_code": "glare_over_total"}	1
3	agent	request_recapture	{"failing_metric": "glare_over_text_frac"}	{"hint_box": [201, 1006, 462, 141], "successor_id": "c_002"}	2

2. The retake: glare fixed, a different rule fires

crop_bottom.jpg arrives in the slot the retake request opened. compare_captures reports the glare fixed, but the new framing cut the bottom of the receipt with text at the cut, so a different rule fires on a different metric. Nothing scheduled this second instruction; the second image's measurements chose it.

RETAKE bottom_edge_clipped

“The bottom of the receipt is outside the frame. Step back and include the total line.”

The clause that fired
metric	value	test	threshold
edge_touch_sides	["bottom"]	contains	"bottom"
bottom_band_has_text	true	==	true

Hint box (page pixels, x y w h): [0, 895, 800, 122]

capture c_002, attempt 2, state after the agent: awaiting_retake

The stored trace
seq	actor	action	inputs	outputs	caused_by
1	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "174e787070788680", "nearest_distance": 64}
2	agent	decide_capture	{"glare_over_text_frac": 0.9211}	{"outcome": "retake", "rule_id": "glare_over_total", "reason_code": "glare_over_total"}	1
3	agent	request_recapture	{"failing_metric": "glare_over_text_frac"}	{"hint_box": [201, 1006, 462, 141], "successor_id": "c_002"}	2
4	agent	compare_captures	{"metric": "glare_over_text_frac", "before": 0.9211}	{"after": 0.0, "status": "fixed"}	3
5	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "330e7c3870707058", "nearest_distance": 64}	4
6	agent	decide_capture	{"edge_touch_sides": ["bottom"], "bottom_band_has_text": true}	{"outcome": "retake", "rule_id": "bottom_edge_clipped", "reason_code": "bottom_edge_clipped"}	5
7	agent	request_recapture	{"failing_metric": "edge_touch_sides"}	{"hint_box": [0, 895, 800, 122], "successor_id": "c_003"}	6

3. The second retake: accepted

clean_a.jpg arrives as attempt 3. Every metric is comfortably inside its band, so the agent accepts it. Accepting is the only thing the agent may finish on its own.

ACCEPT accept

“Every measurement is comfortably inside its band.”

No rule above accept matched.

capture c_003, attempt 3, state after the agent: accepted

The stored trace
seq	actor	action	inputs	outputs	caused_by
1	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "174e787070788680", "nearest_distance": 64}
2	agent	decide_capture	{"glare_over_text_frac": 0.9211}	{"outcome": "retake", "rule_id": "glare_over_total", "reason_code": "glare_over_total"}	1
3	agent	request_recapture	{"failing_metric": "glare_over_text_frac"}	{"hint_box": [201, 1006, 462, 141], "successor_id": "c_002"}	2
4	agent	compare_captures	{"metric": "glare_over_text_frac", "before": 0.9211}	{"after": 0.0, "status": "fixed"}	3
5	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "330e7c3870707058", "nearest_distance": 64}	4
6	agent	decide_capture	{"edge_touch_sides": ["bottom"], "bottom_band_has_text": true}	{"outcome": "retake", "rule_id": "bottom_edge_clipped", "reason_code": "bottom_edge_clipped"}	5
7	agent	request_recapture	{"failing_metric": "edge_touch_sides"}	{"hint_box": [0, 895, 800, 122], "successor_id": "c_003"}	6
8	agent	compare_captures	{"metric": "edge_touch_sides", "before": ["bottom"]}	{"after": [], "status": "fixed"}	7
9	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "174e7870785d8680", "nearest_distance": 64}	8
10	agent	decide_capture	{"rule_id": "accept"}	{"outcome": "accept", "rule_id": "accept", "reason_code": "clean"}	9

4. Not a receipt: escalated, and only a person moves it

not_doc.jpg has no document quad, so the agent does not guess: it escalates. While it waits, every agent tool refuses it — the calls below were made against it and these are their real refusals. Then a reviewer rejects it, and the trace records the reviewer as the actor.

ESCALATE not_a_document

“This does not look like a receipt, so a person will take a look.”

The clause that fired
metric	value	test	threshold
quad_found	false	==	false

capture c_001, attempt 1, state after the agent: escalated

StateError: inspect_capture refused: c_001 is 'escalated'; agent tools act only on received or measured captures
StateError: decide_capture refused: c_001 is 'escalated'; agent tools act only on received or measured captures
StateError: request_recapture refused: c_001 is 'escalated'; agent tools act only on received or measured captures
StateError: compare_captures refused: c_001 is 'escalated'; agent tools act only on received or measured captures
StateError: escalate refused: c_001 is 'escalated'; agent tools act only on received or measured captures
PermissionError: 'approve' is human-only; actor was 'agent'

The stored trace
seq	actor	action	inputs	outputs	caused_by
1	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "eae8a4f4d4d0f8ea", "nearest_distance": 64}
2	agent	decide_capture	{"quad_found": false}	{"outcome": "escalate", "rule_id": "not_a_document", "reason_code": "not_a_document"}	1
3	agent	escalate	{"reason": "not_a_document"}	{"review_item": "opened"}	2
4	reviewer	reject	{"note": "checked by hand, not a receipt"}	{"state": "rejected"}	3

5. A suspected duplicate: a person decides

dup_a.jpg is receipt A re-shot at a different pose. Its perceptual hash lands 3 bits from the accepted clean_a.jpg (the threshold is 6), so the agent escalates it; it can never auto-reject or auto-accept a duplicate. Even the reviewer's approve is refused until they say whether it is a duplicate. Here they confirm it is, and it is rejected.

ESCALATE suspected_duplicate

“This looks like a receipt already in the batch. A person will decide.”

The clause that fired
metric	value	test	threshold
nearest_distance	3	<=	6

capture c_003, attempt 1, state after the agent: escalated

approve before resolving: RuntimeError: c_003 is a suspected duplicate; resolve_duplicate before approving

The stored trace
seq	actor	action	inputs	outputs	caused_by
1	agent	inspect_capture	{"image_shape": [1600, 1200]}	{"phash": "174c7070785d0680", "nearest_distance": 3}
2	agent	decide_capture	{"nearest_distance": 3}	{"outcome": "escalate", "rule_id": "suspected_duplicate", "reason_code": "suspected_duplicate"}	1
3	agent	escalate	{"reason": "suspected_duplicate"}	{"review_item": "opened"}	2
4	reviewer	resolve_duplicate	{"is_duplicate": true, "note": "same receipt, re-shot"}	{"duplicate": "confirmed"}	3