io42 lab ARCH-0001 · Phase 3

The fleet
operating system,
once the traces vote

Phase 3 is what IO42 becomes after every deferred component in Rev D has earned its place through evidence. One rule organizes the whole drawing: authority flows downward, evidence flows back up. Nothing here is scheduled — each block ships only when a measured trigger fires.

Stochastic · advisory Deterministic machinery Safety boundary
DrawingARCH-0001
Phase3 · projection
BasisRev D, Part II
StatusFuture development
01 /Proposesintent enters the system

Coordinator agents

Decompose missions into tasks and propose them into the pool with suggested priorities. Their opinion is an input, never a verdict.

Humans & reactive triggers

Ad-hoc requests through chat, plus standing rules — battery < 15% → dock — that inject tasks with no reasoning loop involved.

02 /Schedulesdeterministic arbitration
Task pool Durable, inspectable backlog. Every entry carries provenance, priority class, deadline, and spatial footprint.
Scheduler Scores by policy, not by vibes: w₁·priority + w₂·urgency + w₃·age − w₄·travel − w₅·resource_risk. Hard preemption for safety classes, an aging bound so nothing starves, and physical affinity — pose, battery, payload — as a first-class term.
Matchmaker Assigns an agent template from trace history: per template, task type, and robot class — success rate, duration, bounce rate, token cost. Historical evidence beats self-estimates. Epsilon-greedy exploration keeps new templates hirable.
03 /Preparesthinking stays outside the lease

Intake queue — warm context

A fresh, task-scoped agent is instantiated and subscribes to the robot's read-only state-delta stream — payload format is the resync digest. It plans speculatively while it waits. Reads never require the lease.

Ready queue — mise en place

Entry requires a concrete, pre-validated goal candidate. Nobody touches the stove until prep is done. Grant order may weight context freshness, bounded by aging so slow thinkers still get served.

Lease authority — one robot, one owner, one epoch

Whole-robot lease with two-phase semantics: deliberation heartbeats, goal-progress renewal. Every grant increments a fencing epoch; stale writes are rejected with LEASE_LOST. Reclaim window for the slow, queue service for the dead, anti-squat forfeiture after K reclaims.

Grant-time re-validation of the candidate goal; a bounce returns the agent to intake with its context still warm. The first write after any epoch change is rejected until the state digest has been consumed.

Safety boundary Nothing above this line moves metal
04 /Executesvalidated, deterministic, safety-rated

Validation gate

Every write is checked against the capability manifest: kinematic limits, safe zones, force ceilings, and enforced perception freshness — advisory poses expire.

Execution stack

Nav2, MoveIt2, controllers, drivers. The safe-idle controller owns the robot whenever no valid lease-holder does. Safety sensors are truth; agents are opinion.

Fleet router — the scale-out seam

Discovery, auth, lease brokering, and command routing across N robots, generated from the same capability manifests as robot #1. Deterministic travel-cost auctions run between robots for spatial tasks, never between agents. Multi-robot transactions — two-lease atomicity — unlock here.

05 /Learnsevidence flows back up
episode_id traces correlation engine pattern mining graduation gate rules & behavior trees

One episode_id threads every MCP call, ROS2 goal, rosbag window, validation verdict, and lease event. Mining proposes; the graduation gate — human review plus shadow evaluation — decides. Rules never self-deploy into the safety layer. The same trace tables feed the matchmaker, so the platform gets measurably better at hiring the longer it runs.

06 /Graduatedhow each component earned its place
ComponentTrigger that firedState
Lease queue & contentionconflicts > 1/day, second task source liveGraduated
Intake / ready splitidle-while-leased > 25%, deliberation-attributedGraduated
Scheduler + task poolpending depth > 3 sustainedGraduated
Matchmaker≥ 2 templates, divergent trace metricsGraduated
Fleet routerrobot #2 physically on orderGraduated
Rules mining≥ 500 episodes of a recurring task typeGraduated
Multi-robot transactionsconcrete two-robot task specifiedGraduated

Principles carried forward

Agents propose The safety layer decides Architects propose The trace data decides Bounded macros, never raw joints Perception is advisory Epochs make lease loss a fact Control returns after the world-model does One episode_id threads everything