The smallest representation of knowledge a model can rebuild.

symbol compiles any source into the smallest text from which a panel of frontier models can recover its knowledge — the claims it makes, not the bytes.

L95 is the metric: the fewest characters needed to recover 95% of a source's weighted claims across the panel. Scores count only the lift over what the models already knew — so it measures compression, not memorization.

sources

reached L95

best compression

2.4×

decoder panel

6 +1 held-out

Leaderboard

best compression first · L95 target

source	chars	claims	L95	ratio	recovery	lift	ceiling
quill-lang-stdlib	1,780	29	758	2.4×	96%	+92%	96%
glowpaper-howto	1,692	25	869	1.9×	95%	+92%	91%
keldra-interview	1,646	20	895	1.8×	95%	+92%	90%
zephyr-protocol	1,676	24	957	1.8×	95%	+93%	93%
festival-de-lumora	1,705	13	—	—	92%	+89%	91%
lumen-yield-study	1,576	22	—	—	92%	+82%	90%
marrow-bay-event	1,940	16	—	—	93%	+88%	91%
tarn-rail-timetable	1,737	25	—	—	94%	+90%	93%
vembari-almanac	1,852	20	—	—	93%	+91%	90%

A dash under L95 means even the full source did not yield 95% robust recovery across the panel — the long tail of unique facts is fragile. Ceiling is the recovery the panel gets from the entire source.