Phenotypic plasticity in oil and gas

The objection that should win

Take the strongest form of the case against continuous intake first, because in oil and gas it is not hypothetical. An integrity engineer who wires more sensors onto a platform does not get more truth. She gets more signal, and signal is not truth — it is correlation with truth, sampled through instruments that drift, foul, and lie under specific and predictable conditions. A hydrogen probe reads high not because hydrogen is present but because it has been recalibrated against the wrong reference gas. A cathodic protection rectifier reports steady current while the anode bed beneath it has gone to open circuit. Satellite InSAR shows ground movement across a pipeline right-of-way that is agricultural subsidence, not corrosion-driven soil loss over the line. Continuous monitoring, on this argument, does not reduce exposure to deception. It multiplies it. Every extra channel is another surface a bad sensor, a spoofed SCADA tag, or an operator under production pressure can use to push a false state into the belief system. A model that updates on everything is not more robust than a model that updates rarely. It is more available to whoever controls the inputs.

This is the sea-turtle problem, restated for steel and hydrocarbons. Turtles nest by cue — brightness gradient toward open water — and the cue used to be reliable. Streetlights broke the correlation and the turtles walked inland to die. A pipeline integrity system that treats every telemetry tick as a reason to revise its belief about wall thickness is exposed to the industrial equivalent: a fouled sensor becomes a false all-clear, or worse, a false alarm that trains the engineer to distrust the channel that would have caught the real event. The objection, stated at full strength, says: you have not solved the reliability problem by adding intake. You have handed it more places to fail.

What survives the objection

Most of it survives, and it should be said plainly rather than argued around. Plasticity has a cost in biology — measured, not asserted. Daphnia that maintain the sensory and developmental machinery for helmet formation pay a fitness cost in kairomone-free water, growing slower than clones that never bothered. The equivalent in an asset system is real: every additional sensor is a capital cost, a calibration schedule, a failure mode of its own, and a training burden on the person who has to interpret it. On a stable, low-variance asset — a buried, cathodically protected, coated pipeline in a geologically quiet onshore field, inspected for forty years without a leak — the canalised strategy wins outright. Monthly aggregation, quarterly pigging runs, five-year direct assessment: fixed intervals, fixed thresholds, cheap. Wiring that line for continuous acoustic emission monitoring is plasticity nobody needed, paid for against a threat that was never going to arrive on an hourly clock.

And the deception risk is not overstated. Sensor networks on ageing infrastructure genuinely do drift, and an integrity system built to revise belief on every reading, without discipline about which readings deserve trust, will indeed be walked into false confidence by exactly the mechanism the objection describes. This is the honest cost of the third rung. It does not disappear because the architecture is elegant.

Where fixed intake actually fails

The concession does not extend to the case that matters, which is the case the industry keeps rediscovering the hard way. The characteristic failure is specific: an integrity signal aggregated monthly, feeding a decision system whose underlying failure mode develops in hours. Hydrogen-induced cracking under a wet H2S environment on a subsea riser is not a monthly phenomenon. Pressure transients from a closed valve slam, or a compressor trip propagating a surge down forty kilometres of line, resolve in minutes and can initiate a fatigue crack in an existing flaw well inside a single reporting cycle. A monthly-aggregated corrosion rate, however carefully computed, is a genotype-level readout — competence fixed at the last retraining of the model, blind to everything that happened between cutoffs. It is the pipeline equivalent of a Large Language Model: a corpus of last month's readings, frozen, deployed as if it still described this month's steel.

A Large World Model equivalent — a bounded-scene system that senses the current episode and adjusts within it — improves matters but does not close the gap. A pigging run gives a rich, in-the-moment picture of wall condition along the whole line, the way a locust's cuticle and behaviour reorganise once crowding crosses a threshold within hours. But the pig run ends, the belief it produced does not update again until the next campaign, and the six months between runs are exactly where a hydrogen crack that started as a subcritical flaw becomes a through-wall leak. The episode closes; the organism's memory of it lapses with the tool retrieval.

What continuous intake with provenance actually buys

The answer to the deception objection is not more trust in more channels. It is less undifferentiated trust, applied to more channels, with the source of every belief carried alongside the belief. A wellhead pressure reading, a seismic reprocessing update, a pipeline cathodic protection log, and a regulatory notice about a newly designated high-consequence area are not interchangeable inputs to a single integrity score. Each arrives with a channel, a timestamp, and a decay function appropriate to its own physics. A hydrogen probe reading decays in trust over days if it has not been cross-checked against a coupon retrieval. A pigging-run wall-loss estimate decays over months, roughly matching the corrosion mechanism's own timescale. A regulatory notice does not decay at all until superseded by another notice. This is the distinction that answers the turtle problem directly: a belief tagged with its source can be quarantined the moment that source is shown compromised — the fouled probe recalibrated, the InSAR pixel reassigned to known subsidence — without touching beliefs derived from channels that are still sound. A belief baked into a fixed monthly aggregate carries no such tag. When the aggregation method is later found to have been wrong, there is no selective correction available; the whole model has to be rebuilt from the next cutoff.

The claim that actually holds

None of this licenses the strong reading that continuous intake beats fixed intake as a general law. It plainly does not, on the stable pipeline, in the quiet field, where the cost of maintaining live streams and provenance discipline exceeds the risk being guarded against. Canalisation — the fixed inspection interval, the frozen model, the monthly number — is itself an engineered adaptation to a genuinely slow-moving hazard, and it is often the right one. The claim that survives is narrower than "watch everything, always, and you will be safe." It is a classification claim: responsiveness in an integrity system sorts into exactly three kinds, distinguished by when observation is permitted to revise belief — never within the deployed period, once per bounded episode, or continuously with provenance attached — and the third kind has no successor. There is no fourth architecture beyond "hold a revisable belief for every stream still running, tagged by the channel that produced it, decayed on that channel's own clock." Beyond that, what remains is not a new kind of intake. It is more sensors, cheaper inference, better calibration of trust per channel, and an integrity engineer who no longer has to choose between a number that is current and a number that is honest.