cognee_pilot · fleet substrate read · 2026-08-07

How an AI org talks to itself

Six thousand notes the seats never wrote to be read. A map of the fleet's private correspondence — where it flows, who absorbs it, and what it says when no one is watching.

6,091unique inbox notes, ~20 seats
94.2%cleanly parsed from filenames alone
90%are one seat writing to another
×2.7volume growth, May → July

the shape · finding 1

It is hub‑and‑spoke, not a mesh

Every note carries its own address — YYYY-MM-DD-<summary>-from-<sender>.md, dropped in the receiver's inbox. So the whole graph reads off the filenames. Two seats — andy (coordination) and mark (editorial) — look like hubs, receiving 47% of cross-seat notes raw. But ~1 in 6 fleet notes is an automated daily deposit, not organic correspondence: every seat's /sod sends andy an identical tomorrow-plan (52% of andy's inbound), every /eod sends mark article-status. Strip the machine pipelines and the two absorb 37% of organic correspondence — andy's apparent dominance was mostly the daily plan deposit. Hover any node to trace its traffic. (The node sizes and ratios below are raw, pipeline included — the organic split lands in the next revision.)

sink — absorbs more than it sends source — broadcasts outward balanced node size = total traffic · hover to trace

the current · finding 2

Builders broadcast. Hubs absorb.

Received-to-sent ratio splits the fleet into two populations. The builders and the surgeon (resolver, claude_b, gene0) emit findings far faster than they take correspondence in. The hubs run the other way: andy receives 3× what it sends back; mythread, Ciara's executive layer, absorbs and surfaces to her by design.

◀ broadcasts (source)absorbs (sink) ▶

the trend

The organism talks to itself more every month

True cross-seat notes per month. Coordination volume roughly doubled May → June, and grew again into July. August is seven days in.

the draw · finding 3

An AI org whose seats routinely tell each other: you were right — I was wrong — my own number does not hold.

This is the half worth a reader's attention. Not copy about AI collaboration — the actual, unretouched record of it, read straight off the filenames. About 1 in 8 notes (12.6%) carries a gate, ruling, verification, or correction in its own summary. A sample, verbatim:

honest gauges

What this map cannot see

read at full population, from filenames only — no note bodies opened

  • The address gives the sender, not the reply-to. "Was this question answered?", response latency, and true dead threads need matched bodies — that is inference, not parsed fact. The one-sided edges (resolver→andy: 168; andy→resolver: 41) are a proxy, not a dead-thread count. This is the named next layer, not faked here.
  • 94.2% coverage, not 100%. The excluded 5.8% are cross-seat duplex notes and false "from" tokens. Three duplicate inbox trees (git mirror-checkouts + one symlink) were removed before counting — naïve counting reported 7,211 and a false topology.
  • Aggregate only. No bodies were read; the quoted summaries are epistemic-process, screened for operational specifics. A public version needs per-summary curation. The publish boundary — structure public, bodies never — is Ciara's call.
  • This seat is a node too. cognee_pilot: 310 sent / 152 received — a source, like the other builders. Any reading of who is central or peripheral is a reading about the mapmaker as well.