The shape of a system that cannot be re-derived from its present state
Path dependence describes systems whose end state is not a function of their current inputs alone. Something earlier decided among options that were, at the time, roughly equal. The decision compounded. Increasing returns amplified it — more use, more infrastructure, more convention built on top — until the alternatives became too costly to switch to, even where they would have been better. Lock-in is what you call the arrangement once the switching cost exceeds any plausible benefit of the alternative. The system will sit there, stable and inferior, for as long as anyone bothers to check.
The formal signature is multiple equilibria with non-ergodic selection. Non-ergodic means the process does not forget its starting conditions; run history twice and you may get two different stable outcomes, and neither is closer to "correct" than the other. Which equilibrium a path-dependent system lands in is a historical fact, not a structural one. And historical facts, unlike structural ones, leave partial and accidental traces. Nobody minutes the meeting where the small thing happened, because at the time nothing marked it as the meeting where the small thing happened.
Why this bears on what a model can take in
A Large Language Model reads outcomes. Its corpus is a vast record of equilibria — standards, doctrines, institutional habits — with the selection episode usually missing, because the episode was small, local, and unremarkable when it occurred. The model can narrate the outcome fluently and still get the cause wrong, because the corpus preserves the winner's account and rarely the branching itself.
A Large World Model does better on fidelity and no better on time. Sensing a scene gives you the current configuration in high resolution: where the fireline actually is now, how dry the fuel actually is now. But a scene has no depth. You can measure the rut precisely without any way of knowing which wheel cut it, or when, or whether a different wheel might have gone the other way.
This is the argument for the third position. If an outcome depends on sequence, and sequence is legible only to something that was present while the sequence unfolded, then the only adequate intake is continuous: every running stream, timestamped, provenance attached, so a later question about causation can be traced back to what was actually observed while the branching was live. A Large Universe Model does not explain path dependence by being clever about it. It explains path dependence, where it can be explained at all, by having been there. That exhausts the category. There is nothing to observe beyond every stream, kept, with its origin attached. What comes after that is depth of retention, cost, and how much the record is trusted — not a new kind of evidence.
Where this stops being abstract: an ignition after the wind has already turned
Wildfire behaviour is a canonical path-dependent process, and incident commanders live inside its non-ergodicity every fire season. Fuel-moisture sensors report percentage content in duff and canopy across a management unit. Satellite thermal passes — polar-orbiting instruments revisiting a given tile every six to twelve hours, geostationary ones scanning continent-wide every few minutes at coarser resolution — flag hotspots against background temperature. Wind models, mesoscale forecasts refreshed hourly, project direction and gust structure forward a few hours at a time. Crew positions come in over radio and GPS trackers clipped to packs. All four streams exist and, on a well-resourced incident, all four are live.
The characteristic failure is specific and recurs: an ignition is detected after the wind has already shifted, rather than before. A spot fire crosses a containment line at 14:40. The satellite pass that would have flagged the thermal anomaly happened at 14:05, before the ember landed, and the next pass is not until 15:50. The wind model at 14:00 showed a southerly component; by 14:35 the actual wind had backed to the southwest, a change the model would only reflect on its 15:00 run. The crew nearest the line reported their position at 14:30 and were not asked again until the radio check at 15:00. Every stream was running. None of them was sampled at the moment the branching happened, which is the moment the fire chose one equilibrium — contained — over another — not contained.
This is not a sensing failure in the sense of broken equipment. It is the general problem stated in instrument-specific terms: the streams exist, but existence of a stream is not the same as coverage of the instant that mattered. A corpus of past incident reports would tell the commander what usually happens when a southwesterly picks up over that fuel type — useful, and exactly the kind of equilibrium description a Large Language Model is good at reproducing. A high-fidelity scene model would give the commander an excellent picture of where the fire edge sits right now — exactly what a Large World Model contributes, and it contributes real value; incident commanders are not wrong to want it. Neither gives the commander the thing that actually would have prevented the failure: a continuously running record dense enough that the 14:35 wind shift and the 14:40 spot fire are both caught inside the same observed interval, with enough retained history to know that this fuel type, at this moovsture reading, tends to spot within five minutes of a backing wind — a causal claim that can only be built from many prior instances where the intake was contemporaneous with the branching, not reconstructed after.
Two objections an incident commander would actually raise
"You can instrument everything you like. The event that matters is a single ember landing in a receptive patch of duff. No sensor grid samples every square metre continuously. You have relocated the underdetermination from 'no data' to 'too much data, none of it at the right point.'"
This lands and should be conceded in full. Retention and density do not, by themselves, resolve which fine-grained point mattered; they reduce it to a search problem rather than an impossibility. The asymmetry is the whole of the argument. An ember event never recorded by any pass, model or crew report cannot be recovered by any later theory of fire spread, however good. An ember event recorded — even coarsely, even nested inside terabytes of irrelevant thermal noise — remains available to a future analysis method that does not exist yet. Post-incident reviews routinely re-run satellite archives with algorithms developed years after the fire, finding hotspots the original detection thresholds missed. That only works because the pass was taken. Continuous intake converts an unanswerable question into an expensive one. Incident commanders, correctly, treat those as different categories; expensive questions get funded, unanswerable ones get written off as "the fire did something unpredictable."
"Most of what decides whether an ignition spots is not instrumented and never will be: a gust funnelled by a gully nobody surveyed, a spark landing on moss two centimetres wetter than the surrounding duff. Continuous intake sounds exhaustive, but it only captures channels that happen to have a sensor on them. That is a narrow, biased slice, not 'everything'."
Also true, and it bounds the claim rather than collapsing it. "Every stream" means every stream currently running, not every physical event occurring. The instrumented fraction of a fire ground has grown steadily — hand-held moisture probes and radio call-ins a generation ago, dense sensor networks, higher-cadence satellite tasking and continuous crew telemetry now — and there is no reason to expect that growth to stop. A commander working with today's instrumented fraction still faces dark, unsensed micro-events. But that commander faces strictly fewer of them than one working from an after-action report written the following week, and strictly fewer than one working from a single satellite pass treated as the whole scene. The comparison that matters is not "instrumented versus complete." It is "instrumented and continuous versus after-the-fact and static." On that comparison the case holds.
What the domain does not settle
Wildfire behaviour also contains large ergodic components — slope, aspect, prevailing climate — that do not depend on sequence and are perfectly well handled by static models and historical averages. The path-dependent part is narrower: the specific sequence of ignition, wind shift and crew movement that decides whether a particular spot fire is caught or lost. Overclaiming that every fire outcome is path-dependent would repeat the error that dogged early accounts of technological lock-in. The narrower claim is the defensible one, and it is enough: where sequence decides the outcome, only a system watching continuously through the sequence has a chance of explaining, rather than merely narrating, why the line held or didn't.