One row per round, to show how we improve. Collect it with scripts/speed.sh
and add the printed line to the table.
run_started_at to updated_at) of
the last completed runs on develop, from the GitHub API. A process
measure, not a program benchmark; the runners vary by 1.4-1.9x from day to
day, so read trends over several rounds, not single rows.success / completed runs
(cancelled ones counted in the note).started_at to completed_at) of each
job over the last 5 push runs, skipped jobs left out (scripts/speed_ci.py,
jobs API, no token). Jobs that stop early (a red bootstrap_compiler) count
with their short time, so read it with the green column.origin/develop of the last 24 h whose
sha is the head of a push run / all such commits.docs/tokenspeed.md (target below 50 k).scripts/self_hosting.py, “Nikaia in all”.lower-std instructions (cold / warm): callgrind counts per docs/measuring.md
(cold about 181 G, warm about 1.3 G as stated there). Not collected by the
script by default: under valgrind it ran over 10 minutes on a shared
machine, so the columns stay empty until someone runs it on a quiet one. Not
collected in the 2026-10-24 rows for that reason.| Date | Commit | CI median s | Green / runs | Tokens per moved line | Share % | lower-std cold / warm (instr.) |
|---|---|---|---|---|---|---|
| 2026-10-24 | 28e01193 | 510 | 1/5 (1 cancelled) | 163,867 | 71.9 | - / - |
Baseline row. Missing: the lower-std counts (see Method).
| Date | Commit | Job medians s (n runs) | Commits with a run, 24 h |
|---|---|---|---|
| 2026-10-24 | 28e01193 | arm64 528, bootstrap_compiler 60, changes 18 (5) | 141/142 |
The last 5 runs were red at bootstrap_compiler/arm64, so only three jobs
ran to a result; the other jobs were skipped and have no time yet.
Agent cost (scripts/tokenspeed.py, 2026-10-24): 88 agents, 272 M context tokens read, agents over 120 turns = 47 % of reads; finished check builders cost 8-15 k tokens per moved line (docs/tokenspeed.md, Runs).