Immune memory and affinity maturation in telecommunications

The forecast that broke in April

A regional network planner spends the autumn building an eighteen-month capacity plan for a metro cluster of four hundred cell sites. The inputs are a rolling twelve months of traffic telemetry — physical resource block utilisation, uplink and downlink throughput per sector, busy-hour Erlangs — fitted against a traffic mix dominated by video streaming at a known average bitrate. The plan sizes backhaul upgrades, schedules carrier aggregation licences, and orders small-cell infill for three saturated corridors. It is signed off in November. In March, a widely used social app pushes a client update that switches its default video codec and raises the average session bitrate by roughly sixty per cent, and simultaneously changes its feed algorithm to autoplay video by default rather than on tap. Aggregate uplink demand in the cluster does not move much — uplink was never the constraint — but downlink busy-hour load in three sectors rises thirty-eight per cent within six weeks, with no change in subscriber count and no billing anomaly that would have flagged it through churn analytics.

The fault alarms arrive first, not the forecast. PRB utilisation crosses ninety per cent in the affected sectors, congestion-related call drops rise, and customer complaints tell the story before the dashboards do, because the capacity plan has no mechanism for noticing that the traffic mix it was built on no longer exists. The infill sites ordered in November are still eight months from being live. The plan was correct when it was made and wrong the moment the assumption underneath it moved. Nobody lied to the planner. The corpus of telemetry he built the plan on was simply a fixed slice of a world that kept moving after the slice was taken.

What the plan actually assumed

The failure is not a forecasting error in the ordinary sense — the model of demand growth was reasonable, the historical fit was good. The failure is architectural: the plan treated a snapshot of traffic composition as a durable fact rather than as one observation in a stream that needed to keep being checked against new observations. Between the November sign-off and the March incident there was no process that continuously re-tested the assumption, weighted new telemetry against it, and downgraded confidence in the codec mix as evidence accumulated. The alarms were the first revision, and they arrived as an emergency rather than as an update, because the system had no channel for ordinary disagreement between the plan and the world.

This is the shape of a much older engineering problem, and biology solved it first.

The immune system's answer

Vertebrate adaptive immunity faces the same structural threat continuously: pathogens mutate faster than any fixed defence can anticipate, so a defence built once and left alone is a defence that expires. The solution, formalised in Frank Macfarlane Burnet's 1957 clonal selection theory, was to stop trying to design the right answer in advance. Instead, generate a huge library of candidate receptors by random gene recombination — a mechanism Susumu Tonegawa demonstrated in 1976 — and let contact with the actual pathogen do the selecting. B cells whose receptors bind decently are pulled into germinal centres, where activation-induced cytidine deaminase, characterised through Cesar Milstein's and Tasuku Honjo's work, mutates their variable-region genes at roughly one base change per thousand per division, a rate a million-fold above the genome's ordinary background. Most mutants die. The few whose binding improved are re-selected, mutated again, and re-selected again, over weeks, until affinity for the actual threat has risen by two to three orders of magnitude over the original naive receptor. The library was never finished at any point in that process. It was continuously revised by contact.

That is affinity maturation: not a better initial guess, but a standing mechanism for improving a guess by repeated exposure to the thing it is supposed to match, with the losers discarded and the record of the winners kept, tagged with the encounter that produced them.

Streams instead of snapshots

Telecommunications planning has, historically, worked like a naive immune repertoire: build the model, freeze it, ship the plan, wait for the next full replanning cycle to revise anything. A network that instead treated traffic telemetry, fault alarms, spectrum filings and churn signals as permanently open streams — never closed, each observation dated and weighted, confidence in any given assumption raised or lowered as fresh telemetry arrived — would not have prevented the March codec change. It would have registered the shift in PRB utilisation within hours as a live disagreement with the standing capacity assumption, not as an alarm storm eight months too late to matter for the infill schedule already in motion.

This is where the concept becomes a lineage claim rather than a metaphor. A model of a network that only ever sees a training corpus fixed at some cutoff is functionally a naive immune repertoire: comprehensive at the moment of construction, unable to mutate on contact with anything that happened after. A model that senses a scene while it is present — live PRB counters, an active alarm feed — behaves like innate immunity: fast, local, reactive to what is directly in front of the sensor, and largely forgetful once the scene changes, which is why the November plan could see the November traffic mix perfectly and say nothing useful about March. The position that keeps every stream open, dates each belief, and revises confidence continuously is the adaptive arm restated for infrastructure.

Two objections worth taking seriously

A memory B cell recognises an epitope, not a situation. Calling this a model of the network confuses shape-matching against known traffic signatures with anything resembling understanding of why demand shifted.

This lands. A telemetry system tuned to detect deviation from an established traffic profile is a pattern-matcher, not a strategist. It will flag that downlink load in a sector no longer resembles its own history; it will not tell the planner that a codec update at an app vendor is the cause, and it cannot generalise from a video-codec shift to, say, a spectrum refarming order changing available bandwidth in the same sector. The analogy to immunity is not a claim about reasoning. It is a claim about intake discipline — about a system that cannot afford to be wrong for eight months acquiring, dating and re-weighting evidence as it arrives, rather than at the next scheduled forecast. Immunity's narrowness is instructive precisely because it shows continuous intake and comprehension are separable. What continuous intake buys is currency, not insight. The planner still has to interpret why; the streams only guarantee he finds out that something changed while there is still time to act.

A system built to distrust its own priors and revise on every new signal is exactly the kind of system that produces false-positive storms — alarms cascading off transient congestion, capacity plans thrashing on noisy churn data, engineers paged at 3 a.m. for load spikes that resolve themselves in twenty minutes. That looks less like a defence and not obviously better than a scheduled plan.

Also correct, and it names something real: cytokine storm and autoimmunity are what happens when a continuously revising system loses its threshold discipline, and telecom networks have their own version — alarm floods during a routine maintenance window mistaken for a genuine outage, capacity re-forecasts thrashing because a single large event (a stadium concert, a regional festival) gets treated as a permanent shift in baseline demand. The immune system's own answer to this was not to close the intake back down. It was to add governance on top of open intake: regulatory T cells, thymic negative selection deleting most self-reactive clones before they ever circulate, checkpoint receptors damping an overactive response. The telecom equivalent is hysteresis on alert thresholds, decay weighting so a stadium spike ages out of the baseline within days rather than distorting the annual forecast, and a provenance tag on every anomaly recording whether it originated from a known one-off event before it is allowed to revise the standing capacity model. The fix for a revisable system that revises badly is better calibration of trust and decay, not a return to frozen plans.

Why there is no fourth rung

The planner's November plan and the March incident describe the entire design space on this axis. Freeze the corpus and you get a repertoire that is naive by construction, however large. Sense only the live scene and you get fast, forgetful reaction with no memory across incidents. Or keep every stream running, tag each belief with the telemetry that produced it, and let confidence rise and fall on contact — which is what an adaptive, provenance-bearing intake architecture does, in a vertebrate immune system and, in argument, in a Large Universe Model applied to infrastructure. Vertebrates have had five hundred million years to improve on that arrangement and have only ever tuned it: how fast to mutate, how strictly to select, how long to remember, how hard to punish a false alarm. No fourth category of evidence appeared. A telecommunications network run on the same principle would not become clairvoyant about the next app release. It would simply stop finding out about the world eight months after the world already knew.