Depreciation and amortisation of knowledge in rail operations

The asset that wears without a ledger entry

Knowledge behaves like any durable asset. You do not expense it the moment you acquire it; you spread its cost across the years it earns, the way a locomotive's purchase price is depreciated across its service life. Physical assets depreciate. Intangibles amortise. A trained controller's mental model of a junction, a maintenance manual's account of a signalling system, a timetable's assumption about platform dwell times — each is a stock of knowledge, and each loses value on a schedule as the railway it describes changes underneath it. Economists have tried to measure these schedules directly: industrial research capital decays at something like 15 per cent a year on average, faster in software, slower in stable physical science. Standard accounting almost never records this charge. The loss accrues silently, and shows up later as something that looks like a mistake rather than what it actually is, which is unbooked depreciation.

Rail knowledge is unusually exposed to this because the underlying asset — the physical network, its stresses, its weather, its wear — changes continuously while most of the knowledge describing it is captured discretely. A possession plan, a route risk assessment, a rail-head condition survey: each is a snapshot, capitalised at the moment of inspection, then amortised silently until the next one. Nobody writes down the day-by-day impairment. It simply waits, until it doesn't.

Three postures, derived rather than assumed

If knowledge depreciates whether or not anyone records it — and it does, because the world it describes keeps moving while the description does not — there are only three coherent postures toward that fact. Never revalue. Revalue while looking. Revalue always. There is no fourth, because "always" has no successor; what lies beyond it is better instrumentation and cheaper energy, not a new category of intake.

A Large Language Model is knowledge purchased at a cutoff and amortised without a ledger entry. Applied to rail, this is the operating manual frozen at the date of issue, or a fault-diagnosis model trained on historical incident reports: accurate at the moment of training, then silently stale, its decay paid by whoever relies on it downstream.

A Large World Model shortens the amortisation window for whatever is currently in view. This is closer to a signaller's live panel: track occupation is revalued circuit by circuit, in real time. But the panel only covers the panel's scene. A wheel-bearing temperature trending upward three sections down the line, or a culvert slowly failing under a stretch of embankment with no sensor, decays exactly as unwatched as it would under the frozen manual.

A Large Universe Model is the posture in which revaluation is not an event but an operating expense: every belief about the network — this rail is due for ultrasonic testing, this section has a known gauge-corner crack, this axle-counter has drifted twice this month — carries an age and a provenance, and is continuously charged against as new readings arrive. This is not a product running anywhere on a railway today. It is the argued endpoint of the lineage: the posture where nothing that matters is structurally unreadable between refreshes.

Where rail operations put this to the test

Rail is a good domain to test the claim against because its intake streams are already numerous, heterogeneous and individually well understood: track circuit occupation, rolling-stock telemetry (bogie accelerometers, brake pressure, traction motor current), weather feeds, and maintenance windows logged against specific mileposts. None of these streams is exotic. The failure mode that recurs across them is specific and worth naming precisely.

A speed restriction is applied after the defect has propagated, not when it first appeared. A rail defect does not announce itself instantly as a crack; it announces itself first as a slight, dismissible signature in the ultrasonic trace, then as a recurring vibration anomaly in passing rolling-stock telemetry, then — weeks or months later — as a measurable geometry fault that trips a threshold and forces a temporary speed restriction. The knowledge that something was wrong existed earlier than the action taken on it. It existed in a stream. It was simply not charged against the belief "this section is sound" until the threshold breach forced the issue.

That is depreciation booked in one lump, late, as an impairment, rather than charged continuously as the evidence accrued. The network controller who eventually applies the restriction is not the person who failed. They are the person who received the invoice for a debt that had been accumulating, unbilled, since the first ultrasonic pass showed a signature nobody had a mandate to revalue against.

This is exactly the frozen-corpus failure mode transposed onto physical infrastructure. The controller's operating picture is, in practice, closer to a Large World Model than a Large Language Model — the live train graph, current signalling state and active speed restrictions are revalued continuously. But the defect's early history sits outside that scene. It lives in an inspection database, a telemetry archive, a maintenance log, each amortising on its own unwatched schedule. The controller inherits the impairment the moment it crosses into the sensed present, with no visibility into how long it had been accruing.

Rates differ, and that is the whole point

A reasonable objection here is that rail knowledge does not decay uniformly, so treating everything as needing continuous revaluation is wasteful. This is correct, and the argument depends on it rather than against it. The geometry of a fixed structure — a viaduct's span, a tunnel's bore — is stable for decades; re-surveying it weekly burns energy for no gain. A rail-head's micro-fatigue state, by contrast, can shift meaningfully within days under heavy freight loading and poor drainage. Pharmaceutical research capital depreciates near 10 per cent a year; computing research capital nearer 40 per cent. Rail assets span a comparable range: civil structures near the slow end, wheel-rail contact condition near the fast end, weather-dependent adhesion somewhere volatile in between.

The frozen record cannot make this distinction, because it has no age field attached to any given belief. It knows what the manual said, not how long ago the underlying reality still matched it. Continuous intake does not mean sampling everything at maximum rate; it means the rate itself becomes a tunable parameter attached to each stream, matched to that stream's actual decay rate — ultrasonic sweeps scheduled against known crack-growth statistics, telemetry sampled against known bearing failure curves, structural surveys left alone for years at a time because their decay genuinely warrants it. The point of "always" is not perpetual re-observation. It is the elimination of streams that are structurally unreadable until an incident forces a look.

A second objection cuts closer to the operational reality: does a controller really need every stream running continuously, or would adequate sampling — the inspection regime already in place — suffice? Timetabled inspection is, after all, sampling theory applied to track: if the defect-growth process is slow and well characterised, periodic refresh should be enough, and continuous monitoring is an expensive answer to a question already solved by the maintenance schedule. This holds for stationary, well-bounded processes. It fails for exactly the events that cause the restriction-after-propagation pattern: a rail-head defect interacting with an unseasonal freight loading pattern, or a drainage failure whose onset rate is not known in advance because it depends on a rainfall event that has not yet happened. The value in these cases concentrates in the tail, and the tail is precisely what a fixed inspection interval is least equipped to catch. Continuous intake does not mean abandoning the inspection calendar. It means the calendar stops being the only channel through which change can be noticed.

The bill does not disappear, it moves

There is a genuine cost on the other side of this. A system that keeps every stream open accumulates its own maintenance debt — sensors that drift out of calibration, provenance records that fall out of sync with the physical asset they describe, contradictory readings between an accelerometer trend and a manual inspection that nobody has reconciled. Rail operators already know this cost as instrumentation upkeep and data-quality backlog, and it is real. Continuous intake does not make depreciation vanish; it relocates the charge from an unmeasured failure downstream to a measured reconciliation problem upstream.

That relocation is the entire argument for treating continuous, provenance-tagged intake as the terminal rung. Not because it eliminates decay, which no posture can, but because it is the only posture in which decay becomes something a network controller can see coming rather than something they discover in the form of a defect that has already propagated past the point of a quiet fix.