Postscript · 16 November 2026
What has actually been seen so far.
Borges closed his stories with dated postscripts that confessed what the narrator could not yet prove. This is ours. Speed is measured. Token savings are not yet: that chart stays a placeholder until the A/B benchmark against Python runs.
C reference0.02
Aleph native, fast path0.115
CPython1.12
Aleph native, before fast path1.71
Source: bench/heavy/fib35.aph, median of 3 runs. The interpreter takes 6.4 s and is left out of the scale.
Tokens per finished taskPlaceholder
Same model, same 20 tasks · includes llms.txt and every retryWord count over a folder
Python
[XXX]
Aleph
[XXX]
Parse numbers, report bad lines
Python
[XXX]
Aleph
[XXX]
Records: sort and summarise
Python
[XXX]
Aleph
[XXX]
Bar lengths are illustrative. Real values come from aleph tokens (T-026) and the A/B run.
Chronicle
- 10 Nov 2026The design documentAleph 4 Agents v1.0: vision, language, Tlön, Uqbar and Orbis on paper.
- 12 Nov 2026The names are fixedThe repository opens and the parts take their names from Borges.
- 15 Nov 2026One day of the loopAn agent writes the core compiler and the native backend, task by task, with review.
- 16 Nov 2026The fast pathUnboxed native code makes fib(35) 15 times faster, with identical traps.
15×faster native code after the unboxed fast path
288tests, including interpreter vs native, byte for byte
< 2ktokens for the complete language in llms.txt
[XX]%fewer tokens than Python per task (placeholder)