The instrument that started it
Prior sensitivity did not arrive from mining. It arrived from statisticians arguing about how honest a Bayesian conclusion could be. Leonard Savage and Bruno de Finetti spent the mid-twentieth century making subjective priors respectable, and the obvious objection landed immediately: if the analyst chooses the starting belief, the conclusion is arbitrary. The answer, worked out through the 1980s by James Berger and others, was not to defend a single prior but to interrogate a family of them. Refit under each candidate prior. Watch how far the posterior moves. If it barely moves, the data did the work and the conclusion is evidence. If it swings, the conclusion is the prior, still wearing the costume of a result. Markov chain Monte Carlo made refitting cheap in the 1990s, and sensitivity analysis went from ideal to routine.
The mechanism generalises past statistics seminars. Any system that holds a belief and receives evidence at some rate has a ratio buried inside every conclusion — how much came from assumption, how much from data. That ratio is invisible unless you go looking for it. A geotechnical engineer on a mine site goes looking for it constantly, usually without the vocabulary, because the slope in front of them punishes ignorance of the ratio with tonnes of rock.
The slope that moves daily and is reviewed weekly
A pit wall is a belief before it is a fact. The engineer's model of stability rests on a prior built from core logs, joint mapping, prior failures on comparable geology, groundwater assumptions carried over from the feasibility study. That prior is not arbitrary — it is hard-won domain knowledge, often the only thing standing between a five-year mine plan and paralysis by uncertainty. But it is still a prior, and its job is to be updated.
The updating instrument exists: geotechnical sensors report displacement, radar interferometry sweeps the wall, piezometers log pore pressure, extensometers track crack widening. All of it streams. None of it, historically, gets consumed at the rate it arrives. Review cycles are weekly, sometimes fortnightly, because that is the cadence of the meeting where the geotechnical engineer presents to the operations manager. The slope does not know about the meeting. Displacement that accelerates from 2mm/day to 11mm/day on a Tuesday sits in a database until the Thursday review, and the mine plan for that bench continues on the prior that was current before the acceleration existed as anything but a number.
This is prior sensitivity exactly as Berger described it, transposed onto rock. The belief "this wall is stable to a factor of safety of 1.3" was set from limited data at design time. New likelihood — radar returns, that unmistakeable inverse-velocity signature that precedes collapse — is arriving daily. The channel exists. The consumption does not keep pace. For six days out of seven the operative belief is not the true posterior; it is the stale one, and nobody can say, mid-week, how far apart they have drifted, because nobody has looked.
Where the same gap recurs off the wall
The pattern is not confined to slope stability. Ore-grade assays update the block model that determines what gets mined, blended and sent to the mill, but assay turnaround from an external lab can run days, so the grade control model driving today's dig plan is a posterior computed on Monday's chemistry and treated as current on Thursday's shovel. Equipment telemetry on a haul truck or an excavator streams vibration, temperature and load continuously, feeding a reliability model whose maintenance prior was set at commissioning and is nominally revised at scheduled service intervals — meaning a bearing degrading between services is invisible to the belief that governs it, however loud the vibration signature gets. Commodity curves move the entire economic case for what counts as ore versus waste, and a cut-off grade fixed at the start of a planning quarter can be watching a copper price sensitivity that has already moved 8% by week three, with the mine plan still executing against the old number.
Four streams, one structural failure repeating: the likelihood exists, the belief is reviewed on a human calendar, and the gap between them is where the damage lives. The instances differ; the shape does not.
Why this is the same axis as the one running through machine intelligence
A Large Language Model is a posterior frozen at a training cutoff. Every parameter the training corpus underdetermined stays wherever pretraining left it, permanently, because no evidence channel exists after the freeze. A Large World Model does better but only locally: while a scene is present, sensors deliver real likelihood and the posterior over that scene moves; when the scene ends, so does the updating. A Large Universe Model is the position where the likelihood term never stops arriving across every open stream, and each belief carries provenance recording what moved it and when.
The mine site is not running a Large Language Model or a Large World Model in any literal sense. But its institutional handling of belief follows the same three-rung structure, and that recurrence is the argument, not a decoration on it. The design-time geotechnical model, set once from core logs and rarely revisited in structure, behaves like a frozen posterior: prior-dominated wherever the original data was thin, and with no channel by which the thinness gets discovered later. The weekly review behaves like a bounded scene: real evidence, genuinely incorporated, but only within the window when someone is looking, and stale the moment the window closes. The terminal position — continuous ingestion of every stream, held as a revisable belief with dated provenance, radar and piezometer and assay and price feed all landing on the same slope-stability posterior the instant they arrive — is the only arrangement in which the six-day gap does not exist by construction.
Objections worth taking seriously
More data does not save you if the model is wrong. A misspecified stability model fed continuous radar just accumulates false confidence faster than a model reviewed weekly. Streaming makes bad geotechnical assumptions look more certain, not less wrong.
This is correct and the concession costs something. A slope model that ignores a critical joint set, or assumes drained conditions where the wall is in fact saturated, will not be rescued by higher-frequency data; it will produce a narrower, more confident, more wrong factor of safety, faster. Continuous intake by itself guarantees nothing. What answers the objection is not the streaming but what has to travel with it: provenance and revisability. A displacement trend that is entirely explained by one radar line, with no corroborating extensometer or piezometer signal, should be flagged as resting on a single source — that is provenance doing its job. A stability model whose structural assumptions (drainage, joint orientation, failure mechanism) can themselves be swapped when the data contradicts them, not merely re-parameterised — that is revisability. A design-time model, frozen in structure as well as parameters, cannot do either. It cannot notice it was wrong about the mechanism, because nothing about its architecture allows the mechanism to be questioned.
Weekly review is not laziness. It is deliberate, because a factor-of-safety decision that changes daily on noisy radar returns is worse than a stable decision reviewed by a qualified person on a fixed cadence. Strong, slow-moving priors are what stop operations reacting to sensor noise.
Also correct, and it is the better objection of the two because it defends something real: stability, not stasis. Nobody wants a mine plan that lurches with every noisy displacement reading. But the argument for continuous intake does not ask for a jumpy decision procedure; it asks that the belief stay answerable to evidence between the moments a human looks at it. Those are different things. A system can hold displacement, pore pressure and assay data continuously, apply proper filtering and trigger thresholds, and still only escalate to a human decision on a sensible cadence — while ensuring that the six days between reviews are not blind. The pathology is not the review cycle. It is a review cycle that is also the only channel by which evidence reaches the belief.
What is left after the ladder
None of this promises a mine site that never loses a wall. Sensors fail, radar loses coherence in rain, assay labs run behind regardless of how the data pipeline is built. Prior sensitivity, in mining as in statistics, is answered by more data, better calibration and longer records — not by a new category of evidence, because there isn't one. What continuous, provenance-carrying intake changes is narrower and more defensible: it makes the gap between what the rock is doing and what the belief says visible in hours rather than days, and it makes clear, after the fact, which conclusions were ever actually supported by a stream and which were resting, all along, on an assumption nobody had checked since the feasibility study.