The Markov blanket in astronomy

The surface around the dome

A survey astronomer does not observe the sky. She observes a feed. The telescope that matters most in her working day is rarely the one she books time on; it is the one already running all night, every night, dumping difference images into a broker. Her actual sensory organ is that broker, and everything she believes about what is currently exploding, fading or moving in the sky is conditioned on whatever has crossed that broker's interface in the last few minutes.

This is a Markov blanket in the strict, original sense Judea Pearl gave the term in 1988: the minimal set of variables that, once known, screens off everything else. Condition on the blanket state and the rest of the universe becomes irrelevant to the update. For a wide-field survey the blanket has a name and a schema: alert packets, each with a coordinate, a magnitude, a time, a classification probability, and a provenance tag saying which pipeline produced it. Nothing about the transient itself reaches the astronomer directly. Only the blanket does.

What actually crosses

Four distinct streams cross this surface, and they cross at different rates and with different half-lives of usefulness.

• Survey alerts: difference-imaging detections, tens of thousands a night from a single facility, most of them asteroids, artefacts and known variable stars.
• Transient broker classifications: a probabilistic label — supernova, cataclysmic variable, tidal disruption, unknown — attached within minutes by a filtering layer, itself a model with its own blanket and its own decay.
• Spectroscopic follow-up: a slower, deliberate crossing. A human or a queue schedules a spectrograph, and hours to days later a spectrum arrives that can confirm or overturn the broker's guess.
• Archival plates and catalogues: the slowest stream, crossing only when someone queries it, but carrying decades of light curve at once — the closest thing astronomy has to a long memory rather than a live sense.

Each of these is a separate channel with a separate mixing rate. An alert is nearly worthless twelve hours after arrival for a fast nova; the same alert is still fully informative six months later for a slowly evolving active galactic nucleus. There is no single decay rate for "an astronomical observation." There is a decay rate per phenomenon, and the belief state has to carry that rate, not just the observation.

What is held

What the astronomer actually keeps — in a database, a personal spreadsheet, or increasingly a shared broker-side classification store — is not the raw stream. It is a belief state per object: a classification, a confidence, a timestamp, and a source. This is the part easy to skip past and the part that does all the work.

A belief that "this is probably a young core-collapse supernova, magnitude 18.3, as of last night's alert, unconfirmed spectroscopically" is a different epistemic object from a belief that "this is a spectroscopically confirmed Type Ia at z=0.04, as of three nights ago." Both might sit in the same table. Only the provenance and the age tell you which one to trust and for how long. Strip the timestamp and source and you have converted a decaying estimate into something that looks like settled fact — which is exactly the failure mode that loses transients.

What triggers revision

Revision is not continuous attention. It is triggered, and the trigger conditions are where the architecture either works or fails. Three kinds of event force an update to a held belief:

First, a new crossing of the same channel — a fresh alert on a known object, updating magnitude and hence the inferred rise or decline rate. Second, a crossing of a different channel that bears on the same object — a spectrum arriving that either confirms the broker's photometric classification or contradicts it outright, which happens more often than survey astronomers like to admit; photometric classifiers have real, quantified false-positive rates against certain contaminant classes, superluminous supernovae masquerading as tidal disruption events being a standing example. Third, an archival cross-match — someone or something querying old plates and catalogues and discovering the "new" transient sits on top of a previously catalogued variable star, which does not just update the belief, it can retract it entirely.

Each of these revisions has to be written back into the held state with a new timestamp and a new source, overwriting or annotating the old belief rather than silently replacing it. Systems that discard the update history lose the ability to tell a genuinely new source from a re-detection of something already explained.

What the operator sees

At the desk, what actually renders is a ranked list: a triage dashboard scored by classification confidence, rate of change, and time since last observation. The astronomer's real decision each night is not "what is this object" but "which handful of objects, out of thousands crossing the surface, justify the cost of a further crossing" — spending minutes of spectrograph time or an override on a queue-scheduled facility.

This is the point at which the blanket's staleness becomes a scheduling problem, not just an epistemic one. A young Type Ia rises and falls on a timescale of weeks; a kilonova counterpart to a gravitational-wave event fades below spectroscopic reach within days, sometimes hours. If the belief the astronomer is acting on is six hours old rather than twenty minutes old, that may already exceed the object's mixing time. The alert was accurate when it crossed. It is no longer a reliable description of the sky by the time a human reads it, decides, and gets a slot allocated.

What it costs

The characteristic failure of this domain is not a wrong classification. It is a correct classification, correctly triaged, that arrives too late to act on: the transient fades before anyone allocates the telescope. This is a pure intake-latency failure, not a modelling failure. The broker was right. The astronomer agreed. The spectrograph was booked. The object was gone.

The cost structure explains why the field has built its infrastructure entirely around keeping the surface open rather than around improving any single model's accuracy. Spectroscopic time is scarce and expensive to redirect; false positives sent to follow-up burn allocation that a genuine fast transient needed. The whole apparatus — brokers, cross-matching, target-of-opportunity triggers — exists to shorten the gap between a crossing and an action, because closing that gap is the only lever available once classification accuracy has plateaued.

Two objections worth taking seriously

You're smuggling in an agent-boundary reading of the blanket that the mathematics doesn't license. Pearl's construct is a static independence statement about a graph, not a live sensory surface belonging to an astronomer.

The objection is aimed correctly at a real overreach elsewhere in this literature, and it should be conceded there. But nothing above needs the contested, Friston-style extension. A transient's true magnitude is a hidden state; the alert stream is an observation channel; the posterior an astronomer holds relaxes towards a broad prior once observations stop arriving, exactly as a Kalman filter's covariance inflates between updates. That is ordinary state-space reasoning, not a metaphysical claim about where an astronomer's boundary sits. The dead-reckoning drift of an inertial navigation unit between satellite fixes is the same phenomenon with a number attached — roughly 0.8 nautical miles per hour of unbounded error growth. A held transient classification drifts the same way, just without a convenient unit of nautical miles.

Retrieval and cross-matching already give closed systems live sensory access. A frozen classifier plus an archival query is functionally the same as a permanently open surface. The three-generation story is engineering, not category.

This is the sharper of the two objections for astronomy specifically, because archival cross-matching looks exactly like retrieval: a query fired into an old catalogue, an answer returned, discarded once the triage decision is made. The difference is what happens to the answer afterwards. A retrieval-style query answers one question and leaves no trace in durable belief; run the same cross-match again tomorrow and nothing has been remembered. A genuinely open surface writes the cross-match result back into the object's persistent record, with its timestamp and source, so that the next alert on that object inherits the retraction rather than repeating the mistake. The mechanism looks similar. What differs is whether revision accumulates or evaporates.

Where this sits

A survey astronomer's working blanket is already the third position on the intake axis, not the first or second — streams stay open around the clock, with no scene boundary and no training cutoff, and the entire discipline's infrastructure exists to manage that openness rather than to avoid it. What astronomy demonstrates cleanly is that opening the surface permanently does not solve latency; it only relocates the problem to allocation and provenance. The terminal rung on this axis was reached by necessity, long before anyone had a name for it, because a transient does not wait for a training run to finish.