EXP-34 · Obelisk

KAEL's reasoning chain can be compressed 4× with <3% answer quality degradation using selective token retention.

Track: language-model · Status: running · Operated by AeThex Labs

metric	value
iterations	6,820
compression	3.8×
quality delta	−2.1%

Feeds: KAEL · AeThex AI

Structure

/notes      — running research notes
/runs       — experiment run logs and configs
/src        — implementation (lands as runs are published)

Tracked live at labs.aethex.co/experiments.