Lineage

Same pattern. New substrate.

NASA flew this architecture on Deep Space 1 in 1999. Plan, Execute, Diagnose — under hard energy and latency budgets, verified by conflict-directed search, with anytime semantics throughout. We picked up where they left off.

1998 · Remote Agent

An AI that flew a spacecraft.

The Remote Agent commanded the engines on Deep Space 1, identified its own faults, replanned around them, and reported back with full provenance for every decision. The architecture had three pillars with clean interfaces — that’s why it survived contact with reality.

Muscettola, Nayak, Pell, Williams · Artificial Intelligence 103, 1998

Remote Agent three-pillar architecture with re-route feedback EUROPA / HSTS Planner Scheduler EXEC / RAPS Smart Executive Livingstone Mode ID & Recovery recovery action (bounded re-route)
Conflict-directed search — minimal explanation from observed mismatch model expected behavior observations sensor stream conflict set components that disagree minimal explanation smallest fix that resolves don’t adjudicate every component — let conflicts direct the search

1996 · Livingstone

Verify by conflict, not per-claim guess.

Livingstone watched observations against a model of how a working system should behave, located the conflicts, and used those conflicts to direct the search for what was wrong. The minimal explanation became the recovery plan.

Williams & Nayak · AAAI-96

2012 → · AEGIS / AutoNav

Autonomy only where latency demands it.

AEGIS added autonomous science targeting on ChemCam for Curiosity. AutoNav drove Perseverance. Both held the discipline of the original Remote Agent: structured action with explicit fallback, decisions reported with provenance, energy budgeted before commitment. Perception added only where the light-speed delay made human-in-the-loop impossible.

Francis, Estlin, Doran et al. · Science Robotics 2(7), 2017

When to add perception — bounded by latency tolerance human-in-the-loop light-speed delay autonomy threshold command + verify perception + plan add capability only when the round-trip cost forces it

The mapping

Three pillars. On a Rust substrate.

Pillar NASA Remote Agent Math-Ground AI
Plan Planner / Scheduler (EUROPA, HSTS) mgai-algebra · BestAction · NarrativeObjective
Execute Smart Executive (EXEC / RAPS) verity-cascade · joule-runtime · ask-core::cascade
Diagnose Mode ID & Recovery (Livingstone) mgai-evidence · mgai-epistemic · joule-compose-correctness

Today · The substrate

Same shape. Four pieces.

The architecture lives today in four substrate pieces — a deterministic cascade that emits its trace as the chain of thought, four modality encoders that share a hypervector space, a reasoning ⊕ generation split per cascade tier, and a model tier that lazy- loads only when the deterministic spine cannot place a request. Each piece carries its own picojoule receipt. The whole architecture map is at /lane.

Open the architecture → See the receipts

Continuation

Same shape. Generalized from a spacecraft to a substrate.

Walk the substrate →