Autonomous mission intelligence.
The category we are building toward, the evidence it stands on, and the reference mission that forces the whole problem.
Autonomous mission intelligence is the software layer that lets a mission composed of many autonomous components, planners, controllers, learned perception, fault management, behave as one governed system.
It is defined by five properties, not by any specific technique: authority (human intent is preserved), uncertainty handling (belief is first-class), resource feasibility (constraints are joint), safety (nothing runs without a permit), and evidence (every decision is replayable).
It is a systems layer, not a model. When we talk about it, we lowercase it, as if it were already an industry term.
Evidence first, architecture second.
Every position on the site traces to something we can point to: NASA's ranked technology shortfalls, JPL flight heritage from Remote Agent to Ingenuity, defense solicitations, and agency roadmaps from four continents. We collected that evidence before we drew the architecture, and we keep collecting it after.
NASA STMD shortfall rankings. JPL mission autonomy papers. DARPA and DIU solicitations. ESA and JAXA roadmaps. ISRO and iDEX release notes.
Multi-agent research with adversarial verification. Claims survive only if three independent votes confirm them against the primary source.
Every architectural claim is a falsifiable hypothesis. When evidence refutes a position, the position moves, not the evidence.
A long-duration planetary surface campaign.
We designed the runtime against the scenario that forces the whole problem: a multi-year surface campaign with a rover, an aerial scout, and a set of distributed instruments; intermittent Earth communications on a light-second scale; compound failures; a science plan that changes as the mission finds what it finds.
Every subsystem the runtime touches, belief state, plan repair, coordination under intermittent links, learned perception with operating-domain contracts, the assurance kernel that survives a radiation event, has to work in that scenario. If it works there, defense satellites and commercial constellations become an easier problem, not a harder one.
Systems designed against slogans get built around edge cases nobody stress-tests. Systems designed against a specific mission get stress-tested from the first week. A reference mission is a discipline, not a customer.
We publish research notes with a fixed structure, position, evidence, mechanism, open questions, references, at /writing. Every essay carries a named human author, a date, and a bibliography drawn from the primary sources. There is no publishing cadence promise. We prefer a quiet month to a filler post.