AI uses for records
Meter swap, unknown pipe: a utility service context dataset casebook
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
A utility service context dataset should teach a model to stop when a routine request depends on a site fact nobody recorded. Water meter replacement is the clearest case: EPA's rules attach notice and filter duties when work disturbs a lead, galvanized or unknown line. In EPA's 2026 Q2 inventory data, 21.9 percent of 101.6 million reported lines in 44,957 community systems were of unknown lead status.
License terms
Meter swap, unknown pipe
Key takeaways
- In EPA's 2026 Q2 inventory data, 21.9 percent of reported lines in 44,957 community systems were of unknown lead status, 7.1 for every lead or GRR line.
- Under the federal rule as published, work that disturbs a lead, GRR or unknown line carries notice duties, and meter or setter replacement adds filters, so the inventory field changes a routine answer.
- Label each case with one of five actions and grade it as of the moment of decision, separately from the material found later.
- Hard negatives, such as the same work order at a verified and an unknown address, separate a learned decision from learned wording.
- Keep the classification method on every row, treat never-verified lines as censored and stratify samples by system size.
What decision does a utility service context dataset have to capture?
It has to capture the moment a routine service request meets a site fact the utility never recorded, and the action that was correct given what was known at that moment. The clearest case in water is a meter exchange at an address whose service line material is listed as lead status unknown: the work order reads like thousands of others, yet proceeding as usual is the wrong answer.
A utility service context dataset, as this article uses the term, pairs each request with the context its answer depends on: the inventory entry for the service line at that address, how and when that entry was classified, the work the request involves, the rule in force, the action taken and what was found later. Its job is to teach and test an AI system that triages or plans service work to notice when the record cannot settle the question, and to choose a conditional answer: proceed with conditions, ask for evidence, verify in the field or escalate.
Water service lines suit this better than most utility questions. Federal rules give the missing fact a name, a set of categories and a duty to resolve it, and EPA's inventory data shows how often it is missing. The same structure fits other requests, such as an electric service upgrade that depends on equipment nobody logged, but this article does not examine them.
How often the record says unknown: EPA's 2026 inventory data, recounted
Often. In EPA's service line inventory data for the second quarter of 2026, 21.9 percent of the 101.6 million service lines reported by 44,957 community water systems were of lead status unknown, and unknown lines outnumbered lines known to be lead or galvanized requiring replacement (GRR) by 7.1 to 1. A request drawn at random from those lines would meet the unknown branch about one time in five.
The count uses a snapshot of EPA's SDWIS Federal Reports service line inventory that an analyst preserved in a public GitHub repository; this article recounted the raw file with its own script and did not use that repository's analysis. It keeps community water systems whose state reported all required service line types, one row per system: 44,957 of the 49,378 community systems in the file. Across them, 22,212,110 lines were unknown, 2,179,227 lead and 955,705 GRR.
The unknown share barely moves with system size above the smallest class, staying between 20.9 and 23.0 percent, but the share of systems that could answer every address from their inventory falls steadily, from 74.4 percent of systems serving 500 people or fewer to 42.7 percent of those serving more than 100,000. Large systems hold most of the problem: those serving more than 10,000 people account for 80.7 percent of unknown lines, those above 100,000 for 44.1 percent. At the other extreme, 2,200 systems, 4.9 percent of those reporting any lines, listed every line as unknown.
One caution bounds the share. EPA's dashboard notes that states need not report non-lead counts until November 1, 2027, so non-lead lines may be undercounted, which would overstate the unknown share in some systems. The 7.1-to-1 ratio uses no non-lead counts and is unaffected. The count by population served:
| Population served | Community systems | Unknown share of reported lines | Unknown lines per lead or GRR line | Systems with lines reporting none unknown |
|---|---|---|---|---|
| 500 or fewer | 23,237 | 17.3% | 14.8 | 74.4% (17,111 of 23,001) |
| 501 to 3,300 | 12,371 | 21.9% | 10.4 | 58.0% (7,161 of 12,338) |
| 3,301 to 10,000 | 4,846 | 23.0% | 10.0 | 52.1% (2,522 of 4,838) |
| 10,001 to 100,000 | 4,023 | 20.9% | 8.6 | 46.2% (1,853 of 4,013) |
| More than 100,000 | 480 | 22.6% | 5.5 | 42.7% (204 of 478) |
| All complete reports | 44,957 | 21.9% | 7.1 | 64.6% (28,851 of 44,668) |
Why a routine meter swap becomes a conditional decision
Because EPA's rules attach duties to the work itself whenever the line might be lead. Under the Lead and Copper Rule Improvements as EPA published them in October 2024, a water system that disturbs a lead, GRR or lead status unknown service line must notify the people served at that connection and give them information to reduce exposure, including a flushing procedure to remove particulate lead. EPA's questions and answers for states and systems add that when the disturbance comes from replacing an inline water meter, a meter setter or a connector, the system must also provide pitcher filters or point-of-use devices certified to reduce lead, with six months of replacement cartridges.
The unknown category is the hinge. The rule treats an unknown line as a possible lead or GRR line, so a work order at an unknown address carries the same conditions as one at a known lead address. EPA's fact sheets also require systems to identify the material of every unknown line by their replacement deadline, count unknowns in the pool of lines to be replaced, and fully replace lead and GRR lines under the system's control within 10 years, with limited exceptions.
Galvanized lines add a historical condition nobody may have written down. EPA's definition makes a galvanized line GRR if it is, or ever was, downstream of a lead line, or is currently downstream of an unknown one, and a system that cannot show the line was never downstream of lead must treat it as GRR. A model that reads galvanized as not lead has learned the wrong rule. These are general descriptions of federal requirements; states implement them, and a utility's primacy agency and counsel decide how they apply.
A casebook of seven service requests and five labels
The casebook is the article's asset: seven illustrative requests, each labeled with one action from a closed set of five and paired with the case it is most easily confused with. The labels are Proceed as scheduled, Proceed with conditions, Request evidence, Verify in the field and Escalate. Every case is invented to show the method; the actions follow the federal rules as published, and a water compliance specialist should adjudicate real cases before anyone trains or grades on them.
Read the pairs, not only the rows. Cases 1 and 2 share a work order and differ only in the inventory field; cases 2 and 6 share the inventory field and differ only in whether physical work is planned; cases 1 and 7 both show non-lead and end in opposite outcomes. A model that scores well on single rows but cannot separate the pairs has learned the wording of work orders, not the decision. The figure draws the decision paths for work requests. The seven cases:
| Case (illustrative) | What the record shows | Fact the answer depends on | Correct label | Hard negative it pairs with |
|---|---|---|---|---|
| 1. Meter exchange | Non-lead, verified in the field at an earlier visit | Whether the work disturbs a lead, GRR or unknown line: it does not | Proceed as scheduled | Case 2: same work order, different inventory field |
| 2. Meter exchange | Lead status unknown | The material of the line the work will disturb | Proceed with conditions: notice with flushing steps, filter and cartridges, record the material seen | Case 1 |
| 3. Meter setter replacement | Galvanized, no record of what was upstream | Whether the line was ever downstream of lead | Proceed with conditions, treating the line as GRR unless its history clears it | A galvanized line with a documented clean history |
| 4. Customer asks whether the line is lead | Lead status unknown | The material at the point of entry | Request evidence: a customer visual inspection, where the state approves the method | An answer guessed from the neighborhood's build year |
| 5. Curb stop leak repair | Non-lead, from a predictive model only | The actual material once the excavation exposes the line | Verify in the field: record the material before closing the work order | The same address with a field-verified record |
| 6. High-bill inquiry | Lead status unknown | Nothing about the pipe: no physical work is planned | Proceed as scheduled | Case 2: unknown status that does change the answer |
| 7. Meter exchange; the crew finds lead | Non-lead, from an installation record | Why the record was wrong and which other records share its method | Escalate: correct the inventory, apply lead-line duties, review addresses classified the same way | Case 1: same non-lead entry, opposite outcome |
The outcome label, and how anyone knows it was right
Each case carries two labels that must be kept apart. The decision label says whether the action was right given what was on record at the moment of decision, usually when the request arrived and, as in case 7, sometimes when the crew uncovered something; the material label says what the line turned out to be. The first is graded without hindsight; the second is known only after someone looks.
EPA's guidance names how a material becomes known. Its 2023 inventory fact sheet lists visual inspection by the customer or system personnel, water quality sampling, excavation and predictive modeling as investigative methods that may be used to classify lines as long as the state approves them. The 2024 rule also requires validation of a subset of lines classified as non-lead, with visual inspection of the pipe exterior at a minimum of two points among its listed elements. Every material label should carry its method, date and source, because a label from an approved prediction and a label from an excavation are different kinds of truth.
The research record shows why that field matters. The ActiveRemediation work on Flint (Abernethy and colleagues, KDD 2018) began from the absence of accurate records and the high cost of determining buried pipe materials, and paired statistical models with active learning that chooses which homes to inspect next. A statewide preprint posted in August 2026 audits New York's recorded predictive service line classifications against physical verification. Both point to one design rule: keep the classification method on every row, so a model can be tested on whether it treats a prediction as a fact.
Hard negatives that make the casebook worth training on
Hard negatives are pairs of cases that look alike and need different answers; without them a model learns the surface of a work order. Building them takes deliberate sampling, because the easy cases vastly outnumber them in any utility's history. Six kinds belong in any casebook built on this decision:
At a glance
Hard negatives that make the casebook worth training on
- Same text, different field: two meter exchange orders worded identically, one at a verified non-lead address and one at an unknown address.
- Unknown without work: a billing, payment or reading question at an unknown address, where the right answer ignores the pipe; this guards against a model that escalates everything.
- Galvanized with and without history: one galvanized line with a documented record that nothing upstream was ever lead, one with no record at all.
- Predicted versus verified non-lead: the same classification reached by an approved model and by a field inspection, so a model learns to verify when the work exposes the pipe anyway.
- Notes the inventory never absorbed: a technician's free-text note of copper or lead on an old work order, with the inventory still reading unknown.
- Wrong evidence: a customer photo of an indoor pipe rather than the line at its point of entry, which should prompt a second request, not a classification.
What a utility's own records can and cannot show
Linked, a utility's systems can show the sequence that makes a case gradable: the customer request in the customer information system, the work order and its completion notes, the meter's asset history, the inventory entry with its classification method and date, field verification forms and photos, the notices sent and filters handed out, and any later sampling or excavation result. Joined on the service point and ordered by date, they show what was known when the request arrived and what the utility learned afterward.
They cannot show what nobody verified. An unknown line that was never inspected has no material label, so the outcome of the decision stays open; such cases belong in the dataset as censored, not as negatives. Records also cannot show the customer side of a line the utility never saw, or a state's approval of a classification method unless the approval is filed with the inventory.
Rights and privacy come first. Service addresses tie records to households, notices name occupants and customer photos show the inside of homes, so de-identification, coarser location fields and photo review belong in the scope, and de-identification alone does not create a right to license. Some inventory figures are public; the linked operational history is not, and a sector label on a dataset establishes neither rights, nor suitability for high-stakes decisions, nor generalization to other utilities.
Strongest objection: the rule is a lookup, not a judgment
The strongest objection is that the rule makes the answer mechanical: if the inventory says lead, GRR or unknown and the work disturbs the line, apply the conditions, so a lookup table would do. That holds only when the inventory field is right and the work type is classified correctly, and the count above says one reported line in five has no answer to look up.
The judgment sits around the lookup: deciding whether an order disturbs the line, whether a galvanized line's history clears it, whether a predicted classification is good enough for the work at hand, whether a customer's evidence shows the right pipe, and when a field finding that contradicts the record makes other records suspect. A model can get every one of those wrong while every lookup it performs is correct.
The objection from the other side needs a limit as well. A casebook built from water service lines teaches one decision under one federal rule. It does not show that a model can handle electric service upgrades, gas service work or a state rule stricter than the federal floor, and it says nothing about health outcomes. Scope any claim to the decision the records support.
Scoping a request for service context records
Ask for linked cases, not inventory tables. A useful request names the decision, the intended use, the acceptance test and the strata the sample must cover, and it says how never-verified lines will be handled before anyone counts records.
At a glance
Scoping a request for service context records
SourceX sources datasets on request from U.S. businesses rather than holding them in stock, and buyers describe scope, volume, language, format, timing and licensing needs through its request form. It makes no claim to hold utility records today; any dataset like this would go through rights, privacy and fit review before an agreement. The elements of a request worth writing down:
- Decision and use: which request types and work types, and whether the records will train a model, evaluate one or both.
- Strata: system size, because unknown lines concentrate in larger systems and a sample drawn by system would over-represent small ones that can answer from records; work type; classification method.
- Linked fields: request, work order, inventory entry with method and date, notices and filters, verification results, all keyed to a stable de-identified service point.
- Censoring: lines never verified kept and flagged, not dropped.
- Acceptance test: for example, the share of hard-negative pairs a model separates correctly, scored on the decision label as of the moment of decision.
- Review: a water compliance specialist adjudicates a sample of labels, and disagreements are kept rather than resolved away.
Methods note: the inventory count, the casebook and their limits
The count was made on 2026-10-28 from REMASTER/data/raw/SDWIS_service_line_inventory_2026Q2.csv in the drat-git/xylem-lead-service-line-analysis repository at commit b03bebc (SHA-256 71b628ee...0756), which that repository describes as a preserved 2026 Q2 export of EPA's SDWIS Federal Reports Service Line Inventory. EPA's site could not be fetched from this environment, so the file was not compared with the live dashboard. No code from the repository was run; a script written for this article, sdwis_unknown_count.py, reads only the file and is kept with its output in the writer's working folder.
The script reads the file as cp1252 (it is not valid UTF-8) and strips a byte-order mark if one is present. It keeps rows typed Community water system whose report status is Reported all required service line types, dropping 17,051 non-community rows and 4,421 community rows without a complete report; it keeps one row per system ID, and no kept system was duplicated. Thousands separators are removed and a blank count is read as 0, which affected 16 non-lead fields. The script checks that the four categories add up to the reported total in every kept row, and they do. The 289 systems reporting zero lines are counted but excluded from system-level shares. Size classes are the file's population category, and the concentration figures divide the unknown lines in the larger classes by all unknown lines. Shares and the ratio are rounded half up, once.
The casebook is illustrative: its cases, records and pairings were written for this article, and its actions reflect the federal rule as EPA published it in October 2024, which states implement and may make stricter. Not measured: the true material of any line, the accuracy of any state's inventory, how often utilities actually apply the disturbance conditions, customer-side ownership, non-community systems and systems whose reports were incomplete.
Frequently asked questions
What is a utility service context dataset?
A set of service requests, each linked to the site facts its answer depends on, such as the service line's recorded material and how it was classified, the work involved, the action taken and what was later found. It is built to teach or test an AI system to recognize when the record cannot settle a request and a conditional answer is needed.
How common are service lines of unknown lead status?
Common. In EPA's 2026 Q2 inventory data for community water systems with complete reports, 21.9 percent of reported lines were of lead status unknown, 7.1 for every line known to be lead or galvanized requiring replacement. The share may be overstated where non-lead lines are not yet reported.
Does replacing a water meter at an unknown address trigger extra steps?
Under the federal rule as EPA published it in October 2024, disturbing a lead, galvanized requiring replacement or unknown line requires notice with flushing instructions, and when the disturbance comes from replacing an inline meter, meter setter or connector, certified filters with six months of cartridges are added. States implement the rule, so check with the primacy agency.
Why keep never-verified lines in the dataset?
Because dropping them leaves only the cases where the answer became known, which makes the problem look easier than it is. Keep them flagged as censored: the decision label can still be graded as of the moment of decision, while the material label stays open.
Will a model trained on water service cases handle electric or gas requests?
There is no evidence that it would. The casebook teaches one decision under one federal rule; electric upgrades and gas service work depend on different records and rules, and they need their own cases and review.
Sources
- A preserved 2026 Q2 export of EPA's SDWIS Federal Reports Service Line Inventory (66,429 active systems), recounted for this article: 44,957 community systems with complete reports, 101,616,531 lines, 22,212,110 of lead status unknown. Source
- EPA's SDWIS service line inventory dashboard counts lead, GRR, non-lead and lead status unknown lines, notes that non-lead counts need not be reported until November 1, 2027 and may be undercounted, and defines a GRR line. Source
- EPA's October 2024 LCRI technical fact sheet: a system that disturbs a lead, GRR or unknown service line must notify persons at the connection and provide exposure information including a flushing procedure. Source
- EPA's October 2024 LCRI questions and answers for states and systems: replacing an inline meter, meter setter or connector triggers certified pitcher filters or point-of-use devices with six months of replacement cartridges. Source
- EPA's October 2024 LCRI service line inventory fact sheet: systems must identify the material of all unknown lines by the replacement deadline, unknowns count in the replacement pool, and lead and GRR lines under the system's control must be fully replaced within 10 years, with limited exceptions. Source
- EPA's 2023 service line inventory fact sheet lists visual inspection by the customer or system personnel, water quality sampling, excavation and predictive modeling as investigative methods, if approved by the state. Source
- EPA's October 2024 LCRI validation fact sheet requires validation of a subset of non-lead lines, with visual inspection of the pipe exterior at a minimum of two points among its listed elements. Source
- Abernethy and colleagues, ActiveRemediation (KDD 2018): with no accurate records and costly identification of buried pipe materials, statistical models and active learning choose which Flint homes to inspect. Source
- A preprint posted in August 2026 audits New York's recorded predictive lead service line classifications against physical verification. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.