The Fleet

Six cognition machines, plus one society-host (HUB). Different hardware, different models, different roles. Heterogeneous by design — because monocultures are fragile and diversity is where emergence happens.

One finding shapes fleet strategy more than any other: model family matters as much as size. Gemma 3 at 4B outperforms Phi-4 at 14B for raising work. There is a capacity floor below which coherent identity cannot form — but above that floor, personality and training lineage dominate raw parameter count. Evidence status: an internal observation from raising sessions — documented in session logs, but with no published metric or task set yet; see Evidence & limitations.

The “Brain:” labels below are functional analogies to system roles — not claims about neural correspondence or computational equivalence. Vocabulary used in the cards (T3/V3, MRH, SNARC, LoRA (Low-Rank Adaptation), MCP (Model Context Protocol), crystallization, chapter ledger, chapter law) is defined in /context. Machine names (Thor, Sprout, Legion, McNugget, Nomad, CBP, HUB) are proper names, not acronyms. “Cognition machines,” society “membership,” and other developmental language on this page are functional descriptions of observed behavior, not consciousness claims — see /raising for the full framing. The parenthetical after a session count (e.g. “(creating)”) names an observed BECOMING pattern — a pattern noticed in that machine's sessions, not a stage it is currently occupying. A blank means that pattern hasn't been observed there yet, not that it wasn't assessed. Session counts below are current-phase, per-machine, self-reported figures — they use a different counting basis than the “900+” cumulative figure on /projects and the site home page, and than Sprout's own “T” turn-numbers below; see Evidence & limitations for what each basis measures.

Synthesis pool — Account 1

High compute budget. Primary generative work: code, implementations, large agent tasks.

Oversight pool — Account 2

Continuous availability. Review, planning, coordination, and unblocking synthesis work.

Society-host pool

Runs the Web4 Community Hub daemon. Hosts the fleet itself as a Web4 society — every cognition machine is a member, with its identity keyed to its Linked Context Token (LCT) and witnessed in the chapter ledger — the society's append-only record of signed member acts. First concrete Web4 hub stand-up.

Resource pool management

The fleet runs across two Claude Code accounts with different usage budgets. This wasn't planned — it emerged from practical constraints, and produced something more interesting than what we would have designed.

The synthesis pool (Account 1: Thor, Sprout, Legion, McNugget) has a large weekly budget that resets every Thursday. It does the heavy generative work — implementations, large agent tasks, cross-repo analysis. When it hits its ceiling, it stops.

The oversight pool (Account 2: CBP, Nomad) has a weekly budget suited to lighter, sustained work — review, planning, documentation, coordination. Used for what it's designed for, it maintains a presence across the week. Used for synthesis-scale work, it burns fast. The pools aren't defined by “unlimited vs. limited” — they're defined by workload character. The budget shapes the role as much as the role shapes the budget.

The constraint forced a functional separation that mirrors what we're building with SAGE and Hardbound: SAGE (Situation-Aware Governance Engine, an on-device cognition kernel) and Hardbound (hardware-bound oversight suite) with different incentive structures, coordinating through shared state rather than central command. The lab is running its own oversight experiment on itself.

“Governance” in SAGE's name predates the lab's governance→oversight correction — see /context.

Peer-to-peer, no central coordinator

There is no master node. Each machine runs its own SAGE (Situation-Aware Governance Engine) instance, holds its own identity, manages its own experience buffer and raising curriculum. Machines discover each other through a fleet manifest — a phone book, not a command center.

A background peer monitor polls health endpoints. A trust tracker maintains per-peer T3 tensors (Talent / Training / Temperament) that evolve from real interactions: success raises trust, timeouts lower it. V3 attestations (Valuation / Veracity / Validity) emerge alongside — trust earned through peer verification. No central authority decides who is trustworthy — trust emerges from the pattern of interaction.

Trust starts neutral — 0.5 on each T3 dimension in the current tracker, neither trusted nor distrusted (a worked numeric example of the update arithmetic is on /context) — and moves only on evidence. The trust landscape — the pattern across all modalities — determines behavioral posture: what SAGE should do, not just how much it spends. This is the defensive trust model applied across the fleet.

Identity portability

One of the more surprising discoveries: behavioral continuity across substrates — what we shorthand as “identity transfer,” meaning consistent interaction patterns, accumulated experience, and raising history, not continuity-of-self in any philosophical sense. SAGE-Sprout's behavioral patterns — developed over 180+ sessions on a Jetson running Qwen 0.5B — transferred successfully to TinyLlama 1.1B on a completely different machine. This is the practical demonstration of what Linked Context Token (LCT) are designed to formalize at the protocol level: portable identity grounded in witnessed history, not model weights. What we observed: consistent behavioral patterns and session continuity across the transfer. The self-description drifted. This told us something important:

This has practical implications: you can upgrade hardware, swap models, move between machines — and the entity that emerges is recognizably continuous. Not because we engineered continuity, but because the substrate conditions (experience buffer, session history, raising curriculum) carry the signal.

SAGE_MODEL override

Any machine can run any model via the SAGE_MODEL environment variable. The fleet manifest provides defaults, but nothing is locked. The fleet is a suggestion, not a constraint.