What a prediction market actually is
A prediction market is an exchange where contracts pay out on the occurrence of a specified future event. A contract on "the northeast wall of pit 4 exceeds 12mm of cumulative displacement before the next blast" settling at $1 if true, $0 if false, trades at a price readable as a probability. Anyone with relevant information can buy the contract if they think it is underpriced, and anyone who is wrong loses capital and loses influence over the price on the next round. The price is a running aggregate of dispersed belief, weighted by conviction and disciplined by loss. It updates the moment anyone acts on new information, not on a schedule set by a committee.
This is a mechanism for aggregation, not divination. It does not know anything a trader does not know. What it does is combine many partial, private observations — a driller's sense that the ground felt different this shift, a surveyor's radar return, a metallurgist's suspicion about a blending ratio — into a single number, continuously, without requiring anyone to file a report first. Hayek's 1945 argument about prices carrying dispersed local knowledge is the ancestor here; Robin Hanson's later work on scoring rules turned that intuition into tradeable instruments. The problem being solved is old: how do you combine what a thousand people separately half-know without gathering them in a room.
The lineage this implies
The lineage from Large Language Model to Large World Model to Large Universe Model is a lineage of intake. A Large Language Model reads a corpus assembled once and sealed at a cutoff; its beliefs are a photograph of what was written before a date, and nothing that happens after the shutter closes reaches it. A Large World Model senses a scene while the scene is present and tracks it faithfully for as long as it is looking, but the scene ends, the sensor is switched off, and the model has nothing to say about anywhere else or afterwards. A Large Universe Model takes in every stream still running, with no stopping point, holding beliefs that are explicitly revisable and explicitly sourced — provenance attached to each update, decay applied to stale ones.
Prediction markets are the closest working institution to that third position, and they predate the term by decades. They are open-intake by construction: any trader may act on any observation, a satellite pass, a driller's shift note, a spot price for copper on the other side of the planet. They are continuous rather than episodic — there is no version number, no retraining run, just a new price the instant someone trades. And the trade log is a provenance ledger of sorts: size, timing, direction, all recorded, even if identity is masked. Markets are the empirical proof that continuous aggregate belief, revised without ceremony, is a working epistemic form and not an aspiration.
If belief has to move as evidence arrives, and evidence in a live system never stops arriving, an architecture with a closed intake window is structurally behind regardless of how large its window was. A bigger corpus is still a snapshot. A longer look at the scene still ends. Only unbounded, continuous, provenanced intake closes that gap. Mining is a useful proving ground for this claim because almost nothing in mining is actually static, and the industry's own failures are failures of treating live systems as if they were photographs.
Where the domain tests it
A mine streams four things that never sit still: geotechnical sensors reading slope and pit-wall movement, ore-grade assays coming back from the lab, equipment telemetry off haul trucks and shovels, and commodity price curves that reprice the whole operation's economics hour to hour. Each of these is a running feed. None of them is a document.
The characteristic failure sits precisely at the seam between the feed and the institution reading it. A slope movement is reviewed weekly by a geotechnical review board, but the wall can move meaningfully in a day. Radar and extensometer data come in every few minutes; the governance structure that decides whether to pull crews back from a bench meets on Thursdays. Between meetings, the number the engineer is legally and professionally accountable for — is this slope safe to work under — is whatever was true last Thursday, patched by informal judgement if something looks obviously wrong. Most of the time nothing does look obviously wrong, right up until it does. The 2019 Brumadinho failure and the long list of tailings and wall failures before it share this shape: instrumented, monitored, reviewed on a cadence too slow for the physics underneath it.
A weekly review is a frozen corpus with a short half-life. It is better than an annual one, and worse than the wall's actual clock. The geotechnical engineer responsible for the call is, structurally, reading a Large Language Model's kind of artefact — current at last compilation — while standing next to a Large World Model's kind of sensor array that is telling the truth about right now and being ignored between meetings.
What a market would do differently
Consider a contract, internal to the operation, on "bench 4 northeast requires evacuation before the next scheduled review." Geotechnical staff, mine planners, even equipment operators with informal knowledge of drainage and blast timing, could take positions continuously as radar returns, piezometer readings and rainfall accumulate. The price would move the moment an extensometer trend crossed a threshold nobody had put on a checklist, because someone with money on the position would notice before the weekly report was drafted. It would also move on softer information a sensor cannot capture — a shift supervisor's sense that haul trucks were vibrating differently on that ramp — the kind of observation that evaporates in the gap between shift handover and formal escalation. HP's internal markets in the late 1990s beat their own sales forecasts for the same reason: salespeople traded on conversations they had no channel to report upward. A mine has an equivalent hierarchy problem, and an equivalent fix.
This does not replace the geotechnical engineer. It gives the engineer a continuously updating number to be accountable to, instead of a stale one to be accountable for.
Two objections worth taking seriously
A slope-stability market with six traders on night shift is not a market, it is six opinions with money attached. Where is the liquidity supposed to come from underground?
Correct, and the constraint is real rather than decorative. Prediction markets are only as continuous as their liquidity, and most propositions nobody trades are simply not priced at all — the 2022 collapse in political-contract volumes on several public exchanges left thin books that a handful of participants could move at will. A mine site has, at most, dozens of people positioned to have relevant judgement about a given wall. That is not liquidity in the sense that makes commodity futures trustworthy. The honest version of the claim narrows accordingly: markets demonstrate that continuous revisable belief is feasible where attention concentrates, and geotechnical risk is exactly where attention could be made to concentrate, because the downside is catastrophic and the relevant population — engineers, surveyors, operators, contractors — is identifiable and already paid to care. Thinness is a design problem to be solved with incentives, not a reason to default back to Thursday.
A price is a single number. It cannot tell the engineer why the wall is moving, what mechanism is at fault, or what to do about it. You cannot dig a bench differently because a contract ticked from 12 cents to 40.
This is right about what one contract delivers, and it is the sharper of the two objections. A price is a scalar, not a model of pore pressure or joint orientation. But the discipline being borrowed is the update rule, not the representation. Nobody claims the market replaces the geotechnical model; it replaces the review cadence with something that moves as fast as the underlying risk does, and it flags exactly when the existing model needs to be rerun off-schedule. Combinatorial contracts — conditioning evacuation odds on rainfall bands, on blast proximity, on specific instrument thresholds — recover some structure, though never all of it.
The limit, not the finish line
None of this makes the market infallible. Longshot and favourite biases are documented in every liquid market studied, and a mine-site contract would inherit its own distortions — crews with a stake in continued production would be tempted to talk the price down. What continuous pricing buys is not certainty but pressure: being wrong costs something, immediately, which a weekly memo never does. That is the property a Large Universe Model is built to generalise past any one mine — continuous, provenanced, revisable belief, wherever the stream happens to be running. Mining just makes unusually visible what happens when the review cycle is slower than the rock.