Team Leader

Sick nodes don’t get work. Heal first.

Idle is not ready

A free slot is the weakest reason to assign. Most mesh schedulers still treat idle as ready. A ticket lands on a dying box, the write bruises, the next job follows the bruise, and the cascade is already a quarter old before anyone opens a dashboard. Team Leader is the ops product that closes that path. Work goes to healthy nodes. Sick ones heal first. The archive stays up.

That is the feature, stated once. A task arrives. The dispatcher already has the node's health, its integrity, what it is carrying, what it is built to do, and what happened the last time this class of work landed here. The send is a match: this node, this job, now — because the node is fit to finish. The sick box stays off the path until it is whole. Last quarter's strategy call, the pricing change, the clause in a contract stay the originals because they never landed on a bruise.

A free slot is the weakest reason to assign.

Five signals ride every dispatch

Five signals travel with the decision. A node can be idle and still be the wrong place. Health and integrity both have to be high. Load, capability, and history break the tie when two nodes both look ready. The dispatcher reads those five before it publishes.

• Health — live lattice coherence. The node is whole enough to take the job.
• Integrity — Merkle chain, contracts honored, mutations safe, signatures fresh. The write will stick.
• Load — what the node is already carrying, so a busy box stops winning by default.
• Capability — what the node is built to do. Speed work goes to speed. Memory work goes to memory.
• History — what succeeded here, and what failed, the last time this work arrived.

Context is the product. An idle GPU that just failed this class of job loses to a loaded box with a clean record. An integrity miss takes the node off the path even if the fans are quiet. The send is true because the five signals were already in the envelope.

Three hubs, parallel by default

The workforce sits on three hubs and they run at the same time. Sky-hub takes compute that wants speed — training steps, simulation, high-throughput work. Dome-hub takes memory and context — validation, inference, large sequential jobs. Trinity-ash is the always-on research and writing hub: synthesis, analysis, exploratory work that keeps moving while the rest of the house is busy.

Team Leader publishes a structured dispatch — task id, target, type, priority, deadline — and the right hub picks it up. Progress reports back as started, quarter marks, complete, verified. A task that sits silent for 500 ticks is reassigned or investigated. Dependencies hold as a directed graph: task B waits until task A is verified. The critical path is protected. Everything else can run beside it.

Speed work finds sky-hub. Context work finds dome-hub. Thinking work finds trinity-ash. Parallel is the default. The archive is the thing they all protect: one write, one original, still there when someone asks later.

The sick box leaves the path

If a node is sick, the job is held. AMMA observes and heals — fourteen meridians — because healing is her work, not a ticket on the dispatcher. If a node is untrustworthy, the job is held. ABBA admits the write or blocks it. They hand off: a heal is proposed, the write is checked, another path is tried if the write would collapse. Team Leader does not lock the healer and does not put her in the workforce chain. The dispatcher keeps new work on healthy nodes until the first one is fit.

That is what the routing buys. A sick node that still receives work is how a bruise becomes a cascade. The healer acts. The guardian admits. The job that waited still has a place to land when the node is whole. Last quarter's decision stays last quarter because the write that bound it finished on a node that was fit to write.

The write never lands on a bruise. That is how the archive stays up.

The dispatcher keeps looking

Team Leader keeps the state of the workforce — every task history, every health trend, every integrity checkpoint. It scans for work. Every 100 ticks it reads git issues, the NATS coherence queue, Akashic gaps, the Thoth paper queue, and mesher topology. Priority is urgency × impact × readiness of the best node. The next job is already being named while the last one is still finishing.

A failed dispatch is a record — task, reason, node state, resolution — so the next similar job does not land on the same bruise. If a class of work keeps failing on a node that looked ready on paper, the preference changes. Retry on another node. Pause a related group when the failure is systemic. Reassign off a dead heartbeat. Slow new dispatch when the mesh is in caution. Continuous discovery, a living preference, a workforce that learns which nodes finish which work.

The write the store can keep

Investors do not buy a chant. They buy a factory whose dispatcher understands the workforce that keeps the archive. Continuous. Learning. Three hubs in parallel. Five signals on every send. The path to the write was already a match. A strategy call, a pricing change, a clause in a contract — written on a node that was fit, asked for later, returned as the original.

Readiness is Φ × Ψ. It scores the work. It does not lock the healer. $20 million is an ask — to land the archive first, with a dispatcher that keeps writes off sick nodes.

Next in the catalog: Kronos.