brain-growth snapshot — 2026-04-18
First daily snapshot of the growth metrics surface. Feeds Pod 24's Brain view. Answers Robert's question directly: "it should be making itself better right... how does it show that to us."
Raw row (one line per day) in state/log/growth.ndjson. Script: state/bin/brain/growth.ts. Skill page: state/skills/brain-growth.md.
Today's snapshot
| Metric | Value |
|---|---|
| skills_total | 122 |
| skills_graduated | 25 |
| skills_prose | 95 |
| loader_pitfalls_total | 290 |
| evals_total | 752 |
| evals_last_24h | 410 |
| rolling_mean_100 | 0.935 |
| rolling_mean_1000 | 0.735 |
| open_breakage_p0 | 0 |
| open_breakage_p1 | 7 |
| parity.claude | 0.688 |
| parity.codex | 0.083 |
| parity.gemini | 0.000 |
| parity.openclaw | 0.238 |
| sync_drift_count_7d | 6 |
| recurring_areas_7d | 81 |
| memory_files_total | 96 |
Headline reading: the eval flywheel is healthy — rolling mean over the last 100 runs sits at 0.935, and the 1000-run trailing mean is 0.735 (older runs pre-date a lot of the shape-gates so the longer window is expected to trail the shorter one). 25 of 122 skills are graduated (have real sidecar scripts); the other 95 are still prose, which is the correct ratio for a system this young. 290 pitfalls captured in loader sidecars is the running tally of "lessons learned that a future agent will not be allowed to forget."
What got better since last snapshot
This is the first row in growth.ndjson, so there is no prior snapshot to diff against. The deltas table below shows today matched against today for both the 7d and 30d baselines, which is the correct degenerate behavior.
Starting tomorrow, the cron at 04:00 local will produce the first real 1-day delta, and every subsequent run will populate the 7d and 30d baseline columns for real.
In lieu of a proper delta, here is the current state of the three signal streams that the "what got better" section will cite going forward:
New skills added in the last 7 days (diff vs 7d-ago): Not computable until a 7-day-old snapshot exists. The skill count today is 122 (the state/index.md header currently claims 123 because a sibling pod's skill has been registered in the catalog but not yet committed to the filesystem at snapshot time). Pod 24 will reconcile the drift.
Loader pitfalls appended (top 3 newest lessons learned): The fixer's "self-write" commit convention (git log --grep "fixer: self-write") returns nothing because the repo currently has one squashed commit labeled initial. Using mtime on state/skills/*.agents.md as the proxy until the fixer accumulates real history:
state/skills/evolve.agents.md— last modified 2026-04-18 18:00state/skills/snappy-resume.agents.md— last modified 2026-04-18 15:16state/skills/snappy-pause.agents.md— last modified 2026-04-18 15:16
Once the fixer pod starts writing its own attribution line, this section will cite the exact lesson text instead of the filename.
Parity climbs (per-runtime delta): Also not computable today. Current levels:
| Runtime | 24h mean | Meaning |
|---|---|---|
| claude | 0.692 | Agentic, hooks wired; the baseline to match. |
| codex | 0.083 | Wired but scoring near zero — the parallel ~/.codex/hooks.json registers but the runtime fails the parity probe. Investigation in flight. |
| gemini | 0.000 | Context-only (GEMINI.md), no execution hook. Expected zero. |
| openclaw | 0.238 | Wired-but-scoring-low — Snappy plugin initialized but reads legacy kernel path instead of the snappy-os loader path. Fix in flight. |
When codex and openclaw climb toward 0.5+, the "parity climbs" line in the next snapshot's wiki page will say so explicitly. That climb is the literal visible answer to Robert's question: it's the brain getting better at operating more runtimes.
How to read future snapshots
Each day the script runs. Each row lands in state/log/growth.ndjson. Each new wiki page (state/wiki/brain-growth-YYYY-MM-DD.md) will have three sections:
- Today's snapshot — the same table as above, new numbers.
- Deltas — today vs 7d-ago vs 30d-ago vs expected direction
(UP / DN / -), rendered by growth.ts --deltas.
- What got better since last snapshot — three concrete cites:
new skills, new pitfalls (lessons), parity deltas.
Pod 24's Brain view reads the ndjson directly, not these wiki pages. The wiki pages are the human-legible version for the weekly Ray update and for anyone skimming the repo to see what the system has learned.
_Authored by Pod 40 on 2026-04-18. Files landed first in commit 5dc2794 ("brain: showcase …") because the evolve / showcase harness swept parallel-pod files in-flight — the same git add -A bug that commit 8f15651 was already fixing. Attribution commit follows._