A field guide to synthetic minds

A fleet of agents, verified against truth.

A self-hosted multi-agent system that routes across model failure families and verifies every decision against executable ground truth. Each verified decision becomes proprietary data.

Running our own operations today, as a self-hosted multi-agent OS.

Read the thesis
A field-guide ink drawing of a cockatiel on a branch, its outline flowing into fine constellation and circuit lines with a single gold thread: the natural and the computational drawn as one line.

how to read this guide

notation
§ marks a section of the argument. Fig. marks a plate. SP-0X marks a specimen.
status
live means running our own operations today. beta means running our own operations, still under test. coming means designed, not yet running. Nothing else is claimed.
term
A governed decision is one that was routed across model failure families, criticized by rival models, and checked against executable ground truth before it shipped. Every status on this site is scoped by that vocabulary.
§ 02

How a decision becomes governed.

Every model carries one failure mechanism, baked into its attention design, so one answer inherits it, and consensus only multiplies the same error. The value is in what happens next: each answer is routed across rival failure families, then proven against executable ground truth. The three figures below state it as a thesis (the problem, the method, the defense), kept honest about what is measured and what is not.

Fig. 01 the ceiling
accuracy number of correlated models achievable space passive consensus saturates here ceiling one model = one mechanism
x · number of correlated models y · accuracy solid line: passive consensus saturates here gold dashed line: ceiling point on the curve: one model = one mechanism tinted region: achievable space
Fig. 01. The ceiling. You cannot vote your way to correctness. More correlated models yield diminishing effective votes; the achievable space is far higher.

Stated qualitatively. Every frontier model is one point on the error-correlation curve with exactly one physical failure mechanism baked into its attention design.

Fig. 02 the four mechanisms
01 SELECTION drops the relevant token 02 COMPRESSION loses the detail 03 DECAY forgets the early context 04 DILUTION averages away the signal
01 selection · drops the relevant token 02 compression · loses the detail 03 decay · forgets the early context 04 dilution · averages away the signal
Fig. 02. Four mechanisms. Each attention design fails at a different point. A critic drawn from a different family is positioned, by construction, to catch the builder’s blind spot.

One builder authors each artifact, never fused. A context-isolated critic from a different family tries to refute it. Disagreements resolve by a concrete refutation or by running it, never by voting.

Fig. 03 verification as offset defense
Fig. 03

Verification as offset defense

Four mechanisms, each failing in a different place. A hazard that slips one panel’s gap is caught by the next, because its gaps sit elsewhere. Drag the control to align the gaps, and watch the rare correlated failure thread through.

hazard ship 01 SELECTION 02 COMPRESSION 03 DECAY 04 DILUTION α β γ
  • 01selectiondrops
  • 02compressionloses
  • 03decayforgets
  • 04dilutionaverages
hazard caught at a panel’s solid bar the rare path that threads all four

On the safety and security posture: Responsible AI, Security.

The same machinery that runs tasks today logs the records on its own live traffic, every decision and model call traced end to end and inspectable. Re-estimating the routing from those records, so the system improves as models drift and new ones arrive, is learned-orchestrator training, gated on compute.

§ 03

The purpose of this guide.

Every decision becomes training data.

Every governed decision becomes proprietary action→outcome→correction data that trains our own vertical models.

01 action 02 outcome ◐ 03 correction ◐ → → → improves the model → stops here today · awaiting first outcome

01 action · live

02 outcome · built, under test

03 correction · built, under test

trace · stops here today · awaiting first outcome

Fig. 00: the flywheel. A governed decision is only a log until its action is linked to a real outcome and a real correction. The machinery that does the linking is built and under test; the loop closes on the first real outcome.

abstract

BunBunLabs is a self-hosted AI lab. It runs a system that sends each decision through more than one AI model, has rival models criticize the answer, and checks the result against real-world data before it ships. Several agents run today on the lab's own operations, one of them hardened and the rest in beta; the specialist models this is building toward do not exist yet. Every verified decision is kept as training data for them.

§ 04

Specimens.

A catalog of the agents and products in the fleet: one architecture across many verticals, one specimen hardened and running our operations, several more in beta, and the rest still coming. Scan the index, or open any specimen for its full account.

Open the field guide →

Dated lab notes →

IndexSpecimens
  1. SP-01 Cockatiel Avian sourcing unit live · internal
  2. SP-02 Mochi Cross-family coding studio near-shipping
  3. SP-03 Mentrox Private decision engine coming soon
  4. SP-04 BunBun Generalist maker beta
  5. SP-05 Bushbush Operations & economics beta · read-only
  6. SP-06 Jackie Growth & curation beta · draft-only
  7. SP-07 Linuxa Fleet metrologist beta · shadow
  8. SP-08 Granny Guardian & governance beta

live, running today · beta, in testing · near-shipping · coming

Specimens // BunBunLabs

Featured specimen

Cockatiel · Avian sourcing unit

SP-01 · Avian sourcing unit

Cockatiel.

live · internal

The fleet's working sourcing agent. It finds parts, weighs suppliers, drafts every outreach for sign-off.

Open in the field guide →
§ 05

For every specimen, not only the human ones.

Profits fund animal rescue.

Read the cause →

§ 06

The register.

as of 12 July 2026, set by hand

  • one agent live and five in beta, all running our own operations live
  • two more specimens in build, none yet proven coming
  • the flywheel's machinery runs on live traffic: outcome joiners, corrections, and a quote desk are built and instrumented live
  • decision and model-call observability: decisions and model calls traced and inspectable live
  • decision records: traces captured, not yet complete training triples coming
  • vertical models: none exist coming
  • profit: none; the ledger is empty
  • outside review: none yet coming

the running log →