A law that was not built to explain tides
William Whewell coined the phrase in 1840, in The Philosophy of the Inductive Sciences, to describe something he had watched happen once and thought ought to be named. Newton's law of gravitation had been fitted to the fall of an apple and the swing of a pendulum. It was then found, unaltered, to account for the orbit of the moon, the shape of the tides and the perturbation of one planet by another — facts nobody had used to build the law. Whewell called this jumping-together of independent inductions consilience, from the Latin consilire. His problem was demarcation: separating a law that has discovered something from a hypothesis merely stretched to fit whatever data it was shown. His answer was structural. A genuine law explains classes of fact it was never built for. John Stuart Mill disputed that this carried any special warrant beyond ordinary induction — agreement was just agreement, however it arose. The argument between them has never fully closed, and it matters more, not less, once the facts in question are gathered by machines rather than gentlemen astronomers.
Consilience, on Whewell's reading, is not about how much evidence you have. It is about whether two sources that had no reason to agree, and every opportunity to disagree, came out the same. Ten thermometers calibrated against one reference are one witness wearing ten coats. A biological argument from embryology and a separate argument from geographical distribution, converging on the same evolutionary tree, are two witnesses. The distinguishing fact is independence of the kind of error each could have made.
Where the same structure appears in a water main
A municipal water utility runs several genuinely different streams at once. Turbidity and chlorine residual sensors sit in the distribution network, reporting continuously. Laboratory assays — coliform counts, THM levels, occasional heavy-metal panels — arrive on a slower, discrete schedule, days behind the water they describe. Pressure telemetry across the network tracks hydraulic behaviour: a drop in one zone, a transient at a valve, a signature consistent with a main break or a backflow event. Maintenance logs record what crews actually did — a repair, a flush, a hydrant left cracked open overnight. Four streams, four different physical bases for error: an optical sensor drifts and fouls, a lab assay has its own delay and detection limit, a pressure transducer has its own noise floor, a maintenance record has the errors of human paperwork.
The characteristic failure of the field is not that any one stream lies. It is that the streams are read separately, and the confirmation that would have mattered arrives after the water has already left the plant. A pressure transient consistent with a backflow event, at 2 a.m. in a residential zone, sits in the telemetry log. A slightly anomalous turbidity reading at a nearby node sits in the sensor feed, within tolerance, unremarkable alone. Neither triggers anything by itself. The lab assay that would confirm contamination is run only after a complaint, days later, once symptomatic cases have already been reported to a health department. The utility engineer, reconstructing the event afterwards, finds that the pressure log and the turbidity trace agreed with each other the whole time — nobody had held them together while both were still live. Consilience existed in the data. It was earned nowhere, because nobody was doing the epistemic work of holding two independent streams aligned at the same moment.
This is Whewell's structure exactly, with the stakes changed. The warrant a utility engineer needs — confidence that this reading is a real contamination event, not sensor drift — cannot come from any one stream crossing its own alarm threshold. A turbidity spike alone might be a fouled probe. A pressure transient alone might be a valve operation logged late. What licenses action, ahead of distribution rather than after it, is the coincidence: pressure anomaly, turbidity deviation and a maintenance record showing an open hydrant nearby, all at once, none of them derived from the others.
Why volume of sensors is not the answer
The instinct to solve this by adding sensors runs straight into the objection that most damages consilience as a concept.
Add enough sensors and eventually they all share a calibration standard, an upstream flow model, a common power supply and a common-mode failure. Convergence among instruments that were never free to disagree confers no privileged warrant — it is just correlated error wearing the costume of agreement.
This is the strongest objection in the field and it is correct as stated. Networked SCADA systems in utilities routinely draw multiple sensor types from the same signal conditioning hardware, the same firmware update, the same telemetry backhaul. A firmware bug that biases turbidity readings low, deployed network-wide, will produce beautiful agreement across forty nodes and mean nothing. The 2003 Northeast blackout is the canonical warning from an adjacent domain: operators and their monitoring tools were working from the same stale state estimate, so every instrument they consulted confirmed the same wrong picture, and the shared error propagated a cascading failure that independent, correctly-provenanced readings would have caught.
But the objection argues for provenance discipline, not against continuous multi-stream intake. Independence cannot be evaluated at all — cannot even be asked as a question — unless each stream retains where it came from: which calibration batch, which firmware version, which upstream model, which timestamp. That is not an incidental feature of good record-keeping. It is the precondition Whewell's concept requires and the precondition a frozen corpus discards by construction. A Large Language Model trained on scraped utility incident reports has no mechanism for asking whether two agreeing accounts of a 2015 contamination event trace back to the same wire report or to genuinely separate observers; the provenance was stripped at ingestion, and the agreement it reports may be one witness dressed as ten. The utility engineer's problem is the harder, live version of the same failure: agreement across sensors sharing a calibration chain is not confirmation, and the only fix is retaining calibration lineage per stream, continuously, so that correlated error can be identified rather than assumed away.
Objection: archives converge too, without simultaneity
A second objection observes that plenty of real consilience in the history of science was assembled from archives, at leisure, long after the fact — radiometric dating agreeing with ice-core chronology, biogeography agreeing with embryology. Nobody needed all these streams live at once. Why should a utility need pressure telemetry and lab assays open simultaneously, rather than reconciled afterwards in a report?
The answer turns on whether the object under study is still moving. The age of a rock stratum does not change while you assemble your evidence for it; archival consilience about it is complete and permanent. Water quality in a live distribution network is not that kind of object. A pressure anomaly at 2 a.m. and a lab assay result available on Thursday are consilient about Tuesday's water, which by Thursday has already reached taps. Agreement established after the fact licenses a finding for the incident report. It licenses no action for the population that drank the water in between. Where the target is moving — a distribution network under variable demand, a reservoir subject to runoff, an outbreak spreading through a service area — a coincidence has a shelf life, and its value decays with the age of its slowest contributing stream. Consilience about a moving target has to be re-earned continuously, which means the streams that compose it have to still be running when the question is live, not archived for later reconciliation.
What the recurrence establishes
None of this is a claim that continuous, provenance-tagged, multi-stream intake solves contamination detection. Weighting four streams of different reliability, deciding how much a maintenance-log anomaly should move a probability estimate against a clean turbidity reading, choosing when disagreement between pressure telemetry and lab assay means investigate versus recalibrate — all of that remains hard, and it is hard in the same way whatever the underlying architecture. What the recurrence establishes is narrower and more structural. A frozen corpus of past incident reports can offer breadth of statement but cannot certify that its agreeing sources were ever independent. A bounded sensing window — a full sweep of sensors, assays and logs held together for one shift — can achieve real consilience across modalities, but the warrant expires the moment the shift ends and the window closes. The position that keeps consilience standing rather than episodic is the one where every stream — sensors, assays, telemetry, logs — stays open, tagged with its origin, so that agreement can be re-tested tomorrow and divergence can be dated to the hour it began. That is not a claim about power. It is a claim that the axis of intake runs out there: past that, there is no fourth kind of evidence to admit, only more discipline in weighing the kinds already named.