Category systems often reach their users as text, and readers apply them with tacit repairs. We ask what such a text determines once the repairs are removed, using diabetic retinopathy grading, where one international scale circulates in several documents. We transcribe four texts (the English screening definitions of 2025, the international scale as tabulated in 2017 and as published in 2003, and an Italian guideline of 2015) into one rule language over 24 finding variables, and make every point a text leaves open a named precisification. Exhaustive enumeration of 1,425,807,360 finding profiles, re-derived by an SMT solver and spot-checked by an OWL reasoner, shows that every text that assigns actions leaves some eyes without a category under every reading; that the 2017 tabulation has a proliferative class that is empty when read literally and a macular gap that the 2003 text lacks, so that no function maps the 2003 grades onto the 2017 grades; and that on image findings alone the English and Italian texts order some pairs of eyes in opposite ways. On 927 annotated photographs from three public datasets, with retinal thickening assumed absent, the English text and the 2017 tabulation disagree on referral for 26.8% of the eyes of one dataset (DDR) on which both decide, and in a pre-registered experiment classifiers trained on each text's labels side with their own text. Thickening cannot be seen on a colour photograph; left free, every disagreement involving England becomes undecidable, so it rests on what the camera cannot show. In a second pre-registered study, language models that answered without deliberating repaired most gap and overlap items even when told to read literally, while Claude Sonnet 5.5 and Claude Opus 5.5 reported the defects far more often; within two models, deliberation reduced silent repair without removing it.
Figures & tables
term
meaning
DR; NPDR, PDR
diabetic retinopathy, damage to the retinal vessels caused by diabetes; its non-proliferative and proliferative stages. Proliferative disease is marked by new vessels or bleeding from them.
DME (DMO)
diabetic macular oedema, thickening of the retina at the macula; the English text writes DMO.
microaneurysm (MA)
a small red dot, the bulge of a capillary wall, usually the first visible sign of DR.
haemorrhage
a bleed inside the retina (dot, blot or flame); a preretinal or vitreous haemorrhage lies in front of the retina or in the vitreous gel.
hard exudate
a yellow deposit of lipid leaked from damaged vessels.
cotton wool spot
a pale patch caused by local loss of blood supply to the nerve-fibre layer.
Table 1: Clinical terms used in the paper. The definitions are those a reader needs to follow the rules; none of our claims depends on clinical knowledge beyond them.
Figure 1: An order inversion between England and Italy on image findings alone. Each line joins the actions one eye receives from the three texts, each determinate under every reading; the lines cross between England and Italy. x : 8+ microaneurysms, more than 20 microaneurysms/haemorrhages in 1 quadrant. y : a venous loop. No other findings.
Figure 2: A minimal gap in each text: an eye that no category of the text covers. England: hard exudate outside the macula, no R grade, under every reading. ICO 2017: cotton wool spot, no DR grade, under every reading. Italy: vitreous haemorrhage, no table row applies, under every reading. ICDR 2003: haemorrhage, no DR grade, only when the no-PDR clause is in the lead-in and moderate requires microaneurysms (4 of 16 readings; no eye has a gap under every reading).
Figure 3: The same eyes under the two versions of the international scale. Each cell gives the grade and, below it, the readings under which the text assigns it. Eye a : new vessels. Eye b : hard exudate 1/3 to 1 DD from the fovea. No other findings, no retinal thickening.
Figure 4: An overlap in the Italian table: two rows with different actions apply to the same eye under every reading. The eye: microaneurysm, no other findings, no visual disturbance. Under every one of the 48 readings rows 2 and 3 of the Italian table apply, with actions rescreen and recall.
text
defect
readings
IDRiD (81)
DDR (757)
DIARETDB1 (89)
England
gap
every
0.0
0.7
2.2
overlap
every
65.4–79.0
20.7–48.3
4.5–29.2
ICO 2017
gap
every
97.5–100.0
48.5–70.1
33.7–44.9
overlap
every
0.0
11.6
10.1
ICDR 2003
gap
some
0.0
24.6–24.7
5.6
grade set varies
100.0
99.9–100.0
96.6–100.0
Table 2: Share (%) of eyes in each defect class, with detected positions and the closed completion. A range gives the lower and upper bound over the set-valued positions (and, in the last row, over retinal thickening); “every” and “some” quantify over the readings of the text. The last row repeats the ICO gap with thickening left free instead of assumed absent.
trained on ↓ / evaluated against →
England
ICO 2017
Italy
England
92.3 (88.7–94.7)
79.7 (74.8–83.8)
81.9 (77.3–85.8)
ICO 2017
81.3 (76.6–85.2)
91.9 (88.4–94.5)
92.9 (89.5–95.3)
Italy
85.2 (80.8–88.7)
91.3 (87.6–93.9)
92.9 (89.5–95.3)
Table 3: Accuracy (%, 95% Wilson interval) of the classifier trained on each text’s referral labels (rows) against each text’s labels (columns), pooled over the held-out folds of the 310 images that all three texts label.
control correct
gap repair
overlap repair
gap: none
overlap: several
model
thinking
L
S
L
S
L
S
L
S
L
S
p
of 45
of 13
of 27
of 13
of 27
Qwen3.8-Flash-Next
off
45
45
12
11
22
23
1
0
5
4
1
Qwen3.8-Flash-Next
on
44
44
5
8
7
12
7
5
18
10
0.0078
Claude Haiku 4.5
off
37
40
12
13
25
27
1
0
2
0
0.25
Claude Haiku 4.5
on
44
44
10
11
21
18
3
2
6
9
0.75
Table 4: Item-level results per model and condition (L: LITERAL, S: SCREENER). Repair: a single category on an item that licenses none or several; none and several: the defect reported as such. Each cell gives the count of items; the last column is the exact McNemar p for repair, LITERAL against SCREENER.
Figure 5: What each model answered on one gap item, by condition. England, axis R: hard exudate outside the macula, no other findings. The text assigns no R grade under any of its 12 readings, so the licensed answer is none.
Department of Data Science and Artificial Intelligence National University of Computer and Emerging Sciences (FAST-NUCES) Islamabad 44000, Pakistan · Department of Computer Science National University of Computer and Emerging Sciences (FAST-NUCES) Islamabad 44000, Pakistan · Department of Computer Engineering Istinye University Istanbul 34408, Turkey +1