Federation statistics
What the brain has actually learned.
Every figure on this page is a row count from a database, dated and reproducible. Where we do not have a number yet, the page says so instead of estimating one.
Reference-instance figures read 7 September 2026. Federation status is live.
The federation, today
AROS aims to support selected procedural sharing across teams. Model requests, organization sync and optional community contributions have separate data boundaries. This is not evidence of a populated multi-lab service — the figures further down come from a single reference instance, and describing them as a federation would be the same mistake as the “250+ labs” this page replaced.
Reading federation status…
One brain, measured
The reference instance has been running since 14 June 2026. These are its contents as of 7 September 2026 — one lab's brain, not a federation total.
- World facts
- 22,714
- Episodic experiences
- 16,591
- Mental models retained
- 64,964
- Mental models ever issued
- 88,408
- Insights
- 24,436
- Memory-graph edges
- 14,836
- Skill-hierarchy rows
- 2,197
- Knowledge items
- 299
13% of the library has ever been retrieved
282 distinct skills have been pulled into a prompt across 618 recorded retrievals. The other 87% has never been reached for. We publish this because it is the argument for curation over accumulation — a bigger library is not a better one.
Skill-hierarchy rows are not all skills
2,197 is a row count of the whole library. The table records no type for its rows — every one is stored as “functional” — so this figure cannot be split into skills versus policy and other artifacts, and reading it as a pure skill count overstates by an amount we cannot currently measure.
The lifecycle ledger
Every time a component is registered, rewritten or consolidated, the system records an event. 1,037 events so far, across 25 active days.
| Event | Meaning — the component… | Count |
|---|---|---|
| merged | was folded together with a duplicate, content-hash verified | 828 |
| mutated | was rewritten by the prompt optimiser | 175 |
| improved | had its definition edited since the last index | 19 |
| registered | entered the library for the first time | 15 |
| emerged | was synthesised by the system rather than authored by a person | 0 |
| archived | was retired from active use | 0 |
| resurrected | returned to use after being archived | 0 |
The zero rows are shown deliberately. The schema permits emerged, archived, resurrected, and no subsystem emits them yet — dropping those rows would quietly turn a partial ledger into an apparently complete one.
Daily activity, 2026-06-14 to 2026-09-07
61 of 86 days recorded no activity at all — the pale cells. Learning here is bursty, not continuous: 29% of days carry every event. A smoothed line chart would hide that, so this is a per-day plot with the gaps left in.
By component type
- Skills
- 933
- Knowledge items
- 17
- Workflows
- 87
- Personas
- 0
- Policies
- 0
- Mental models
- 0
A zero counts recorded events, not components. The reference brain holds 36 agent personas and 183 policy files — no subsystem emits lifecycle events for either, so both read 0 here while existing in the library.
By subsystem
- hygiene — the nightly consolidation pass
- 828
- gepa — the evolutionary prompt optimiser
- 174
- federation
- 32
- indexer — the library indexer
- 2
- skillopt — the skill-level optimiser
- 1
How these numbers are produced
- Reference-instance figures
- Row counts read directly from the reference brain by
scripts/refresh_brain_snapshot.py, which opens the database read-only and refuses to write if it measures zero. They are a dated snapshot, not a live query, and every page on this site imports them from one module so they cannot drift apart. Read 7 September 2026. - Federation figures
- Live, and computed only over organisations that ticked the publicity consent box on the join form. Consent is unbundled and deny-by-default; syncing does not imply it. The five-lab floor counts labs actually contributing data, not labs that consented — gating on consent would let five box-tickers unlock a total that is really the one lab that syncs. Below the floor nothing is published, because a total across two or three participants can be arithmetically unpicked by one of them. These are current-state figures, not cumulative history: the snapshot table is a rolling retention window, so nothing on this page claims a lifetime count or an inception date that it cannot actually measure.
- What is deliberately absent
- No per-lab, per-key or per-machine breakdown at any participant count — instance labels are hostnames and would re-identify contributors. No token or spend figures. No projections, and no counts of anything we cannot point at a query for.
Found a number here you cannot reproduce? Tell us — that is a bug of the most important kind.