The conversation a climate scientist is always having
A climate scientist reading a regional dashboard is not reading a report. She is holding up her end of a conversation that has been running for decades and will not stop while she is on holiday. The satellite passes overhead every few hours with a new swath of radiance. The station network reports hourly means, some late, some flagged, some quietly decommissioned without anyone updating the metadata. The buoy arrays drift, lose contact, resume contact from a slightly different position. A reanalysis product folds all of it together on a schedule and asserts, with a confidence interval, what the atmosphere and ocean were doing last Tuesday. None of this is background. It is the other party in an exchange, and she is expected to keep her model of it aligned closely enough to act.
This is the right frame for asking what continuous intake actually buys a monitoring system, and where it does not help at all.
Common ground, briefly
Herbert Clark's account of grounding treats a conversation as a running ledger of propositions both parties accept as mutually held. Every utterance proposes an addition. Every acknowledgement, correction or silence disposes of the proposal. Robert Stalnaker had already named the ledger the common ground in the 1970s; Clark and Marshall showed in 1981 that trying to define it by mutual knowledge alone collapses into an infinite regress — I know that you know that I know — and that real interlocutors dodge the regress with copresence heuristics instead of infinite certainty. Clark and Schaefer, then Clark and Brennan in 1991, turned this into grounding proper: a collaborative process with visible costs, repairs, and least-effort trade-offs. The whole apparatus exists to solve a narrow problem — how two parties manage to refer to the same thing without swapping their entire mental contents.
Applied to climate monitoring, the ledger is the working consensus that a basin is behaving as expected, a station's record is trustworthy, a threshold has not been approached. The grounding acts are the ones a scientist performs constantly: cross-checking a satellite anomaly against a buoy reading, querying a station that has gone quiet, retracting a trend estimate when a reanalysis product is revised upstream.
Position one: the loop only needs to close on request
There is a serious case that continuous intake is beside the point, because grounding is bounded by purpose. Clark's own criterion says parties align only as far as the current task requires. A scientist studying Atlantic meridional overturning does not need live grounding on Southern Ocean sea-ice extent; she needs it when she asks. A frozen archive of reanalysis output, refreshed on a fixed release cycle and queried with a long enough context window, reproduces most of what feels like continuous awareness. The rest is theatre.
"You don't need the ocean to be watching you back. You need the record to answer your question when you ask it. Everything else is instrumentation for its own sake."
This is not a weak objection. Most climate science runs perfectly well on batched, versioned products — ERA5 updated monthly, station networks reconciled on a quality-control schedule, satellite Level-3 composites released in fixed windows. Nobody is demanding that every buoy ping be grounded into a live conversation the instant it arrives. The grounding criterion does real work here: it is exactly why climate archives are batched rather than streamed to begin with. Effort is expensive; ground is built only as far as the task demands.
Position two: purpose itself moves, and the record cannot follow if it is frozen
The counter-case is not that more grounding is always better. It is that the purpose changes when the world changes, and a batched record cannot notice that the purpose has changed.
A marine heatwave forming off a coastline nobody was tasked to watch does not announce itself through the release cycle. It shows up as an anomalous buoy reading three days before the monthly reanalysis would flag it, cross-corroborated by a satellite sea-surface-temperature pass that a fixed-schedule product would not ingest until its next run. The threshold gets crossed in the gap between batches. The scientist assigned to the basin next door is grounded perfectly well in her own region and knows nothing about this one, because nobody's task included it. This is the characteristic failure of the whole enterprise: not a bad model, but a well-grounded conversation happening everywhere except where the event is.
A frozen or batch-refreshed system cannot repair this, because repair requires evidence arriving during the exchange, not before it. A long context window full of last month's reanalysis is a longer monologue, not a live update. The purpose-boundedness argument is correct about how much grounding an ordinary query needs; it says nothing about what happens when the fact that would change the purpose was never written down anywhere the system could read it before the window closed.
What continuous intake actually supplies here
Set against a Large Language Model and a Large World Model, the shape of the difference sharpens. A Large Language Model trained on decades of climate literature and reanalysis summaries can discuss overturning circulation fluently and can propose additions to common ground — "this looks consistent with a known decadal pattern" — but it cannot check that proposal against Tuesday's buoy telemetry, because its intake closed at a cutoff. A Large World Model bolted onto a single monitoring station grounds properly while its sensors are live: it sees the buoy's drift, registers a satellite pass overhead, revises its estimate in real time. But its common ground is scoped to that station, that scene, and dissolves when the tasking ends — it was never watching the neighbouring basin and has no mechanism for being told to.
The position that does not have this failure mode by construction is the one where satellite passes, station networks, buoy arrays and reanalysis products are treated as streams that never finish arriving, each belief about basin state carrying who asserted it — which instrument, which pass, which version of the reanalysis — and when, with the standing option to retract when a later pass contradicts an earlier one. That is not a bigger sensor. It is the same grounding loop climate scientists already run informally, made structural rather than personal, so it does not depend on one person's tasking to notice the gap between basins.
| Property | Large Language Model | Large World Model | Large Universe Model |
|---|---|---|---|
| Intake | closed at training cutoff | live, scoped to one scene | open, across all tasked and untasked regions |
| Grounding duration | none — proposes, cannot verify | for the life of the scene | continuous, with decay |
| Provenance | absent | implicit in the sensor | explicit per belief |
| Failure mode | stale confidence | scene ends, ground vanishes | threshold crossed outside tasking |
The two objections that actually bite
The purpose-bounded objection above deserves the concession it got: most climate monitoring needs very little live grounding, and demanding continuous intake everywhere is an expensive answer to a question nobody asked in the routine case. What it does not cover is exactly the boundary-crossing case, where the required ground is not more history but the newest slice of it, arriving from a region that was nobody's job.
The second objection worth taking seriously is about symmetry. Conversation, on Clark's account, is mutual: both parties can be surprised, both can repair, both are accountable for the ledger. A network of satellite passes, buoys and reanalysis feeds is not a participant that can be embarrassed by a wrong assertion. Pile up enough streams and you have built better surveillance, not a better interlocutor — the analogy to conversation quietly stops earning its keep. This is fair, and it should be stated narrowly rather than argued away. Continuous intake is a precondition for grounding, not a substitute for the accountability that makes grounding matter. When a threshold is crossed in an unwatched region, the open question afterwards is not whether the data arrived — with continuous intake it did — but who was responsible for treating it as an assertion requiring a response, and whether that person had any way to know the assertion had been made. That is a question about institutional readback, not sensing. Tenerife's tower and cockpit both had the transmission; neither had confirmed closure on what it meant. Provenance without an assigned reader is a ledger nobody is checking.
What narrows, not what resolves
Neither position wins outright. The grounding criterion is correct that most of climate monitoring's actual workload needs bounded, on-demand alignment, and a system that tried to ground everything continuously against everything would be spending effort nobody asked it to spend. The counter-case is correct that the residual — the threshold crossed off-schedule, in the gap between tasking — is exactly where batched intake fails structurally and cannot be patched by a longer window. The honest resolution is not that continuous intake wins the whole argument. It is that continuous intake settles the evidence question — every stream, held with provenance, revisable — and leaves untouched the older, harder question grounding was always about: who is listening for the assertion nobody was assigned to hear.