Large Language Thing

Home/Concepts/Translation and untranslatability in municipal water systems

Translation and untranslatability in municipal water systems

Faithful carriage of meaning requires continuous access to the referent, not a snapshot of it. This follows from what fixes reference: not the definition, but the ongoing practice…

Two readings of "safe"

A word like "compliant" sits on every inspection sheet a utility engineer signs. It looks stable. It is not. Its meaning is fixed by a standard, and the standard moves. The Lead and Copper Rule Improvements lowered the lead action level to 10 parts per billion; PFAS maximum contaminant levels were set for six compounds in a rule that did not exist a decade earlier. A well certified "compliant" in 2015 under the old threshold and never retested carries the word forward with none of its original backing. The term is a shape. The content has been swapped out underneath it, and nothing in the file says so.

This is a translation problem before it is a chemistry problem. The engineer reading last year's log is translating from a past state of the language of compliance into the present one, and the two states do not agree on what "safe" denotes. Set two positions against each other on what fixes that agreement, and neither collapses cleanly.

Position one: continuous intake closes the gap

The first position holds that a system streaming sensor arrays, assay results, pressure telemetry and maintenance logs without interruption can keep every governed term correctly anchored, because it never stops checking the referent against the world. Chlorine residual, turbidity in NTU, backflow preventer certification status, cross-connection inspection date — each is a live quantity, not a stored fact. If the standard for "acceptable turbidity" changes from 1.0 to 0.3 NTU, the system does not silently keep applying the old cutoff to new readings; it records the revision, timestamps it, and flags every asset whose last known reading predates the change as unassessed rather than compliant. The word does not rot because the word is never left alone with an old value. This is the Large Universe Model claim in miniature: denotation as maintained belief, with provenance and a decay clock, so a shift in the standard shows up as a logged event rather than an unnoticed mismatch.

Under this position, the characteristic failure of the domain — a contamination event confirmed only after distribution rather than before — is exactly the failure that standing intake is built to prevent. The event is confirmed late when the term "safe to distribute" was being read off a snapshot: yesterday's assay, last week's calibration, a threshold six months out of date. Continuous re-anchoring moves the confirmation point earlier because it keeps asking what the term denotes right now, against the current stream, rather than trusting what it denoted at the last full test.

Position two: some referents are simply gone

The second position does not dispute that live streaming helps. It disputes that it settles anything. Two objections carry the weight here.

The first is Quine's. Watching a rabbit run past and hearing "gavagai," the field linguist can never rule out that the word means undetached rabbit part rather than rabbit — more observation multiplies data, not decisions. The water-system analogue: a SCADA feed shows pressure and residual for a district metered area, and the engineer infers "the main integrity term applies to this whole zone." But the sensor array samples at fixed points on a network with unmapped branch connections, some laid before 1950 with no surviving material record. Continuous telemetry from the points that exist does not determine what the aggregate term "zone integrity" refers to across the points that were never instrumented. More stream is not a decision procedure. The gap between "this reading" and "the thing the report calls compliant" is logical, not just evidentiary, and no volume of intake dissolves it by itself.

Streaming more sensor data at a system with unmapped joints and undocumented service lines is not closing the interpretive gap. It is generating more precise readings of a network whose boundaries were never fixed in the first place. The uncertainty is structural, and no amount of throughput touches structure.

The second objection is about referents that have actually vanished. A 1920s cast-iron main, replaced piecemeal over decades with no as-built drawings, has a "material composition" term in the asset register that nobody alive can verify by observation. No stream, however continuous, recovers a referent that stopped being recorded before recording began. For this class of case — decommissioned wells, abandoned easements, pre-regulation service line inventories — the third position offers the same nothing the first one does. There is no live feed for a pipe nobody photographed before it was paved over.

Where retrieval fails and provenance does not

A third argument deserves a hearing because it looks like it dissolves the whole dispute. Utilities already run scheduled recalibration: instruments are re-certified annually, standards tables are updated when regulations change, GIS layers are refreshed. Isn't this just maintenance, already solved, no need for a distinct category of system?

It solves drift in value, not drift in identity. Recalibrating a turbidity sensor corrects a number against a fixed definition of turbidity. It does nothing when the definition itself is what changed — when "acceptable" silently moved from 1.0 to 0.3 NTU and an old maintenance log, written before the change, is read today as if it meant the current threshold. A refreshed standards table gives you the new value with no trace that an old value with different implications ever governed the same asset. Historical incident reports get read through present definitions, and a 2016 log entry marked "residual within limits" is treated as equivalent to a 2024 one, when the limits it satisfied and the limits now in force are not the same claim. What is missing is not a bigger table. It is a record that says: this term, this asset, this date, this definition, superseded on this date by that one. That is provenance and versioning, not recalibration.

A refreshed index tells you the current answer; a versioned one tells you that the question changed.

Narrowing the claim

Both positions are right about different halves of the same word. Continuous intake — sensor arrays, assay feeds, pressure telemetry, maintenance logs, held with provenance and revision history — does move the confirmation point for contamination earlier, and it does convert silent mistranslation into a logged, inspectable gap, for every term whose referent is still live somewhere in the network. That is not nothing. It is most of what an engineer actually needs on a working system: knowing that "compliant" now carries last Tuesday's PFAS assay rather than 2015's lead test, and knowing when a term has gone unrefreshed long enough that it should be treated as unknown rather than assumed.

But it does not resolve indeterminacy in Quine's sense, and it does not restore what was never captured. A district metering area boundary inferred from partial instrumentation remains an inference, however dense the readings. A main laid before records were kept stays untranslatable in the strict sense: the word for its composition survives, the object it once pointed to is unrecoverable, and no stream, however wide, retrieves it.

The thesis holds at the scale it was built for. Standing intake with provenance is the terminal move for keeping a term correctly anchored to a referent that still exists and still generates evidence. It is not the terminal move for referents that have already been lost, and it does not manufacture determinacy where the underlying network was ambiguous before any sensor was installed. What continuous intake buys the utility engineer is not certainty about every governed term. It is a working distinction between the terms still tethered to a live stream and the terms that have quietly come loose — which, for a system whose worst failure is finding out about contamination after the water has already gone out, is close to the whole of what matters operationally, and honestly short of everything the word "safe" is sometimes asked to guarantee.

Continue