The pessimistic meta-induction in forestry and wildfire

The pessimistic meta-induction in forestry and wildfire

Here is the objection at full strength. Every belief system that fire science has built has eventually been superseded, sometimes catastrophically. Fuel-load models from the 1970s assumed suppression would keep interior forests wet enough to resist crown fire; decades of accumulated dead wood proved otherwise. The 10 a.m. policy, official U.S. Forest Service doctrine from 1935 to the 1970s, held that every fire should be out by mid-morning the day after detection; it produced the very fuel accumulation that made later fires unstoppable. If forestry's best-supported operational theories keep turning out wrong, then an inductive step across that record says today's fire-behaviour models — rate-of-spread equations, crown-fire initiation thresholds, the whole apparatus an incident commander leans on — are probably wrong too, in ways nobody currently notices. Calling any intake architecture "terminal" looks premature against a discipline with this track record. That is a strong case. It deserves to be taken seriously before it is narrowed.

What the historical record actually shows

The argument is Larry Laudan's, from 1981, aimed originally at scientific realism rather than at fire management. Laudan catalogued theories that were well evidenced, competently held, and false: phlogiston, caloric, the luminiferous ether. His point was that predictive success at the time does not license belief that a theory is approximately true, because the historical base rate of success-then-abandonment is too high to ignore. Forestry supplies its own list without needing to borrow philosophy's. The 10 a.m. policy was well evidenced by the standards of its era — early suppression genuinely reduced acreage burned, for a while. Geosyncline-style confidence about fuel continuity across a landscape assumed static moisture gradients that satellite thermal passes later showed shifting hour to hour. Rothermel's 1972 spread model, still taught, still used, has known blind spots in steep terrain and under extreme wind that show up as underprediction precisely when the stakes are highest. The pattern is not that forestry science is bad. It is that confident, well-supported operational belief keeps getting overtaken by streams nobody was reading closely enough at the time.

The bridge to intake

The relevant question is not whether any given fire-behaviour model is currently correct. It is what a system is structurally permitted to observe, and whether it can find out it was wrong. A model trained once on a fixed corpus of historical fire behaviour holds whatever that corpus made likely at the moment of training. It has no channel through which this afternoon's wind shift can inform it that its spread prediction has failed. A system built around a bounded sensed scene — a single fire ground's current sensor mesh, read once — can update within that scene, but the update does not outlive the incident. Neither arrangement can receive an overturning. The pessimistic meta-induction, read procedurally rather than sceptically, says exactly this: a belief-holding system must be built for supersession, not storage, because supersession is the historical norm, not the exception.

Fireground reality gives this teeth. The characteristic failure mode is well known to every incident commander: an ignition is detected after the wind shifted, rather than before. Fuel-moisture sensors report a stand as within safe thresholds at 0600. Satellite thermal passes, on a six-hour revisit cycle for many public systems, last updated hotspot positions overnight. A wind model run at 0400 assumed a ridge-driven diurnal pattern. None of these streams is wrong on its own terms. The failure is that the belief "this flank is holding" was formed from data that stopped arriving before the wind did, and nothing in the architecture flagged that the belief's grounds had gone stale. The commander who orders crew positions on that belief is not making an error of judgement. They are making a correct decision from an incorrectly current picture.

What survives from the objection

Two counterarguments to the terminal claim carry real weight, and the domain sharpens both.

The first is the structural-realist reply: theories are refined, not simply discarded, and what survives supersession is often the durable mathematical or physical core. Rothermel's equations were not thrown out when their limitations in extreme wind became apparent; they were re-scoped, treated as a limiting case valid in a narrower wind-and-slope band, with newer coupled fire-atmosphere models handling the regime where they fail. That is genuine continuity, the forestry analogue of Fresnel's equations surviving the death of the ether. But continuity is visible only after the fact. In 2018, nobody knew in advance which parts of the standard spread model would still be trusted after the next extreme-wind event and which would be quietly retired. A live incident command system cannot pre-sort its beliefs into the durable core and the disposable interpretation before the fire tells it which is which. It has to hold the whole model revisably, spread rate and wind-coupling assumption alike, and let the discovered failure determine what gets kept. That is the architectural point, and it stands whether or not the eventual refinement preserves 90% of the original model or 10%.

The second is the engineering objection, and it is the sharper one for this domain. Continuous intake is not free. A fireground with every stream running — fuel-moisture telemetry, thermal passes, wind model updates, GPS crew positions, radio traffic — produces contradiction constantly: a sensor reading that disagrees with a satellite pass by twenty minutes' latency, a wind model that revises its own forecast mid-shift. An always-updating picture with no way to tell drift from correction can be worse than a stale one, because a commander who trusts every update equally is chasing noise rather than fire. This is a real cost, and it is why the third position in this lineage is defined by provenance, not merely by volume of intake. A belief about flank stability has to carry its grounds: which sensor, what timestamp, which wind model run, how many minutes since verification. When one of those grounds is discredited — the fuel-moisture sensor turns out to have been reading a burned-over patch, the wind model's boundary condition was stale — every downstream belief built on it has to be findable and downgraded, not silently left standing. Without that bookkeeping, continuous intake is a liability. With it, contradiction becomes information rather than noise: a disagreement between two live streams is itself a signal that one of them has failed, and the system can say which claims depend on the failed one.

The narrower claim that holds

Strip away the overreach and what remains is specific. A frozen corpus of past fire seasons cannot learn that this wind shift falsified its spread assumption; its error grows with time since the corpus closed and it cannot report that growth. A bounded sensed scene, however good its sensors, loses everything at the end of the incident and starts the next fire from nothing. The arrangement that survives the objections is the one with every stream still running — fuel moisture, thermal, wind, crew position — held as revisable belief with provenance attached, so that when a sensor or a model run is discredited, the chain of decisions resting on it can be traced and marked. That is not omniscience about fire behaviour, and it does not claim the wind shift will be caught in time. Sampling is still selective; bandwidth is still finite; an incident commander still works from an incomplete picture, always. What the arrangement buys is the capacity to know, after the flank was lost, exactly which reading was too old to trust and which decision depended on it — and to carry that grounds forward into the next incident rather than starting again from a frozen best guess. There is no fourth kind of evidence beyond continuous access, honestly tagged with its own expiry. Getting better from here is a matter of more streams, faster revisit cycles, sharper provenance — scale and trust, not a new category of intake.