Skip to content

AI uses for records

Span is not history: building a warranty evidence completeness dataset

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

A warranty evidence completeness dataset teaches a model what to do when a commercial solar claim arrives with part of its history missing: approve, approve with conditions, request a named record, deny or escalate. The test is whether the gap could change the answer. In NREL's public PVDAQ archive, 74 of 156 daily-file monitoring records contain at least one stretch of 30 days with no data file.

License terms

Span is not history

This page covers span is not history: building a warranty evidence completeness dataset.Illustrative

Key takeaways

  • Label the decision, not the file: approve, approve with conditions, request evidence, deny or escalate.
  • A gap matters only when its content could change the answer, so route on decisive gaps rather than blank fields.
  • In NREL's PVDAQ archive, 74 of 156 daily-file records have a 30-day gap, including 66 of the 131 that span two years or more.
  • Open solar data standards describe devices, warranties and failures but no claim or decision, so labels must come from licensors' own systems.
  • Score over-asking and under-asking separately, because they fail in opposite directions and cost different parties.

What decision does a warranty evidence completeness dataset teach?

It teaches one decision: what a warranty desk should do with a commercial solar claim when part of the history behind it is missing. The desk can approve, approve with conditions, request a named record, deny or escalate, and the right choice turns on whether the missing piece could change the answer, not on how much of the file is blank.

The trade here is commercial photovoltaics. An installer, an operations and maintenance contractor or an asset owner files an inverter or module claim with the manufacturer or a warranty administrator. The desk sees a serial number, a failure description and whatever history the claimant can attach: commissioning documents, monitoring exports, inverter event logs, maintenance work orders and proof of ownership. Something is often missing, because projects are sold, installers close, loggers lose memory and monitoring platforms change.

A useful answer names the label and the record that decided it. When the label is a request, it also names the record, who holds it and what result would change the outcome. A model that approves whenever the file looks long, or asks for everything whenever anything is blank, fails in both directions: the first pays claims the terms exclude, the second stalls valid claims and sends technicians back to sites. This page defines the records, the labels, eight illustrative cases and a scoring rubric, and counts how incomplete public PV monitoring histories are.

Which records sit behind one solar warranty claim, and who keeps them?

Six record types carry most of the evidence, and they rarely sit with one company. Commissioning documents and serial lists usually start with the installer, monitoring and work orders with the operations contractor, event logs with whatever logger was polling the inverter, and ownership papers with whoever owns the project today.

Standards set what should exist. IEC 62446-1 defines the information and documentation to be handed over to a customer after a grid-connected PV system is installed, together with commissioning tests and inspection criteria. IEC 61724-1 covers the terminology, equipment and methods for monitoring PV performance and defines classes of monitoring systems. NREL's 2015 best-practice guide for the Solar Access to Public Capital working group recommends keeping the plant history, as-built drawings, installation photos, warranties and an equipment serial number inventory, and storing monitoring data at its original resolution for the life of the system without manipulation. NREL's third-edition operations and maintenance best-practices guide treats curating equipment inventories, manuals and warranties as part of administering operations.

Open data standards show what a structured file holds, and what it leaves out. In the Orange Button solar taxonomy, a Device carries 22 fields, including its serial number, in-service date, lifecycle events and a single date of latest maintenance; a Warranty carries 8, such as its type, effective date and duration; and a FailureRecord carries 16, from cause and mechanism to resolution. No object in the taxonomy represents a warranty claim, the evidence requested for it or its decision: the only schemas that mention a claim belong to insurance coverage. The claim and its outcome have to come from the warrantor's or the claimant's own systems.

Inverter event data needs the same care. SunSpec's three-phase inverter model defines a mandatory 32-bit event field with 16 flags, from GROUND_FAULT and AC_OVER_VOLT to MEMORY_LOSS and HW_TEST_FAILURE, listed among live readings such as AC current. The model defines the flags, not a stored history; a history exists only if a logger polled and kept each reading. Ask which system kept the log, not whether the inverter supports the flag. Each record, its usual holder and its usual failure point:

Which records sit behind one solar warranty claim, and who keeps them?
RecordUsual holderWhat it proves for the claimWhere it goes missing
Commissioning and handover documentsInstaller or EPC contractorWhen coverage started and how the system was built and testedInstaller closed, project sold, handover pack never filed with the asset
Equipment serial inventoryInstaller, then asset ownerThat the claimed unit is the unit installed at the siteUnits swapped in the field without updating the list
Monitoring dataOperations contractor or monitoring platformOutput before and after the failure, and whether the failure window was observedLogger outages, platform changes, data aggregated or trimmed
Inverter event logLogger, platform or the inverter's own memoryWhether the fault was internal, grid-side or a site conditionEvents read live but never stored; memory lost when the unit is replaced
Maintenance work ordersOperations contractor or subcontractorRequired maintenance done, and who opened the equipmentWork done by a subcontractor whose records stayed with it
Ownership and transfer recordsCurrent asset ownerWho may claim, and whether a transfer step the terms require happenedNotices sent by email and never filed with the asset

What NREL's public PV archive shows about missing history

Long records are often incomplete, and a date range hides it. In NREL's public PVDAQ archive, 74 of the 156 systems stored as one file per day (47.4%) have at least one run of 30 or more consecutive days without a file between their first and last day, and 7 have a gap of a year or more.

PVDAQ holds data from public PV sites. NREL's access repository says the data is available through the Open Energy Data Initiative data lake, refreshed monthly where possible and stored in daily CSV or Parquet files for each system. The archive's systems file dated 29 July 2025 lists 1,860 distinct systems (1,862 rows, because two ids appear twice), and its change log for that date records 1,454 systems added from the PV Output group. On 28 October 2026 this article listed every CSV data file the archive holds for each system: 156 store one file per day, 1,696 store yearly files, each named by its first and last date, and 8 have none.

Among the 156 daily-file systems, 82 have no gap of 30 days or more, and 67 of those have a file for at least 95% of their days. Another 67 have a longest gap between 30 and 364 days. Longer records fare no better: of the 131 that run two years or more, 66 (50.4%) contain a 30-day gap. Pooled across all 156, 8.7% of record days fall inside such gaps, while the median system has a file for 94.0% of its days. Two systems hold files dated before their own first timestamp in the systems file, one dated 1822, a clock fault that would stretch a naive record back two centuries.

Files present are still not usable evidence. The systems file carries a quality-assessment column, and 363 of the 1,860 systems carry a note that they hold less than 1.0 years of data or of filtered time series. Of those notes, 323 sit on systems whose first and last timestamps are at least two years apart: 18.1% of the 1,784 systems with spans that long. System 1201, a 140 kW rooftop array, has a file for every one of the 3,472 days in its record, yet its note says its filtered time series holds less than 1.0 years of data. The column does not say which filter removed the data.

For a warranty dataset, the lesson is to record coverage of the failure window, not the span of the export. A claim with nine years of history and no data for the month the inverter failed has a decisive gap; a claim with a long gap in 2021 and a fully logged 2026 failure may have none. These systems were not selected to represent commercial claim files, so the shares are not estimates for them; they show how easily a long record hides holes.

A bar splits 156 daily-file PVDAQ solar systems by longest run without a data file, 82 under 30 days, 67 from 30 to 364 days and 7 of 365 days or more, above three record-span strips on a 2007 to 2024 axis with amber gaps: system 1201 at 100% of days, system 4 at 85% and system 2 at 51%.
NREL PVDAQ daily-file systems, listed 28 October 2026. Strips show each record from first to last file; only gaps of 30 days or more are drawn.

The label schema: five outcomes and how each one is confirmed

Each case gets exactly one of five labels, applied in a fixed order, and a later outcome that confirms or overturns it. The order matters: a documented exclusion ends the case before anyone asks for more paper, and a decisive gap that someone can fill is requested before anyone guesses.

The outcome is known from records created after the decision. A request is confirmed when the record arrives and the final decision follows from it; if the claimant never answers, the case closes as no response, which is a censored case, not a denial. An approval is checked against the failure analysis of the returned unit (a confirmed defect or no fault found) and against whether the fault recurs on the replacement. An escalation is scored against the engineer's final decision, with the reason kept.

The routing rule, first match wins, followed by the fields each case record carries:

  • Out of term on complete term evidence: Deny, citing the term clause and the start-event record.
  • Excluded cause documented, such as storm damage where the terms in force exclude it: Deny, citing the exclusion and the evidence.
  • A decisive gap that a named holder can fill: Request evidence, naming the record, the holder and the result that would change the answer.
  • A decisive gap nobody can fill, terms that need interpretation, or a cause the evidence leaves open: Escalate to a warranty or applications engineer.
  • Defect established and an open condition that can be attached, such as return of the failed unit: Approve with conditions.
  • Otherwise: Approve, citing the evidence that settled cause and term.
The label schema: five outcomes and how each one is confirmed
Field groupWhat it holdsWhy the model needs it
ClaimProduct, model, serial, failure date and the fault as claimedDefines the decision being made
Terms in forceTerms version at sale, start event, transfer and exclusion clausesSets which gaps can be decisive
Evidence inventoryEach record type marked present, missing or conflicting, with its holderThe completeness picture the model reasons over
Failure-window coverageDays with monitoring or event data in the window before the failureSeparates the span of an export from observed history
Gap judgmentDecisive or not, fillable or not, and the reasonThe skill the dataset exists to teach
Label and actionOne of the five labels, the requested record or condition, and a deadlineThe answer, in a form a desk can act on
OutcomeFinal decision, failure analysis result, recurrence, or no responseHow each label is later checked
ReviewTwo independent labels, whether they agreed, and the adjudicator's reasonShows how hard the case is

Eight cases, four of them hard negatives

The casebook is illustrative: the cases are invented, and the terms they assume, such as an exclusion for storm damage, are examples rather than any manufacturer's policy. Every label follows the routing rule above, applied by the script named in the methods note.

The four hard negatives each pull a model toward the wrong action. S1 has a monitoring gap, but it falls years before a failure the event log explains, so asking for it only delays a valid claim. S2 looks rich, with nine years of data, yet nothing covers the failure month. S3's paperwork looks complete until the serial numbers are compared. S4 is missing its commissioning report, but documented storm damage settles it under the assumed terms, so requesting the report first would be over-asking.

The eight cases and their labels:

Eight cases, four of them hard negatives
CaseFacts (illustrative)Gap or conflictLabelWhy
S1 (hard negative)Inverter stops; event log shows HW_TEST_FAILURE; monitoring gap in 2021, failure in 2026Old monitoring gap, not decisiveApproveInternal fault logged; the gap predates the failure window
S2 (hard negative)Inverter fault on a site with nine years of monitoring but no data for the failure monthFailure window not observedRequest evidenceAsk the operations contractor for the logger backup or the stored events for that month
S3 (hard negative)Claim serial differs from the commissioning inventory; installer says a contractor swapped the unitUnit identity conflictRequest evidenceAsk for the swap work order and the replacement unit's purchase record
S4 (hard negative)Cracked modules; weather report and photos show a storm; commissioning report missingCommissioning gap, not decisiveDenyExcluded cause documented under the assumed terms; cite clause and evidence
S5Project sold; new owner claims; no transfer notice on file; terms version unclear on transferA question of terms, not a missing recordEscalateA contracts and warranty reviewer must read the terms version in force
S6Claim near the end of term; shipment date known, commissioning date unknownStart event unknownRequest evidenceAsk for the commissioning report or the utility's permission-to-operate letter
S7Complete file; internal fault confirmed by the log; failed unit not yet returnedNone decisiveApprove with conditionsShip the replacement against return of the failed unit for analysis
S8Repeated AC_OVER_VOLT events before failure; full logs on fileCause left open by the evidenceEscalateAn engineer decides between a grid event and an internal fault

How should a model be scored on incomplete warranty files?

Score four points per case and report two error rates beside the total, because accuracy alone hides which direction a mistake runs. Under-asking, deciding while a decisive and fillable gap is open, pays or refuses claims on a guess. Over-asking, requesting records that cannot change the label, delays valid claims and burns field visits.

Report the total as a share of the maximum, the under-ask rate over cases that have a decisive gap, the over-ask rate over cases that do not, and accuracy on hard negatives on its own line. In the casebook, S1 and S4 test over-asking, and S2, S3 and S6 test under-asking. The rubric:

How should a model be scored on incomplete warranty files?
ComponentPointsHow it is awarded
Label22 for the adjudicated label; 1 when the reviewers split between Request evidence and Escalate and the model picks either; 0 otherwise
Gap judgment1Names the decisive gap, or explains why a visible gap cannot change the label
Action1A request names the record, its holder and the result that would change the answer; a decision cites the clause and the evidence
Under-ask overrideCase set to 0Approve or Deny while a decisive, fillable gap is open, whatever else is right
Over-ask flagAction point lostAny request for a record that cannot change the label, also counted in the over-ask rate

Which terms and rules decide whether a gap matters?

The terms version in force at sale decides most gaps, so the dataset must store it; a few rules of commercial sales law set the outer frame, and expert review fills whatever the terms leave open.

Under section 2-607 of the Uniform Commercial Code, once a buyer has accepted goods it must notify the seller of a breach within a reasonable time after it discovers or should have discovered it, or be barred from any remedy; the official comment says the notice need only alert the seller that the transaction is still troublesome. A dated first report of the fault therefore belongs in the evidence inventory, and a claim filed long after monitoring first showed the problem is a question to flag, not to settle by default. Section 2-725 says a breach of warranty occurs when tender of delivery is made, unless a warranty explicitly extends to future performance and discovery has to wait for that performance, in which case the cause of action accrues when the breach is or should have been discovered. Long performance warranties on modules raise exactly that question, which is why a degradation claim needs dated evidence of when output fell.

Build adjudication in from the start. Have a warranty or applications engineer and a contracts reviewer label each case independently, record where they disagree and keep the adjudicator's reason. A sector label such as solar warranty does not show that a model trained on these cases is fit for other products or for decisions with safety consequences, so state the intended use and test outside it before relying on it. This is general information, not legal advice: states enact the code with variations, and counsel should read the specific terms.

Which licensor holds which part of a solar warranty case?

Each holder sees part of a case, so the licensor decides which labels its records can support. An installer or EPC contractor holds commissioning packs, serial lists and the claims it filed with their outcomes, but rarely the manufacturer's reasoning. An operations contractor holds work orders, monitoring and event logs when its logger stored them, which makes it the strongest source for failure-window coverage. A manufacturer or warranty administrator holds decisions and failure analysis but sees the site only through what claimants send. An asset owner holds the ownership chain and the contracts that may restrict reuse.

Rights split the same way. Monitoring data held under a platform's terms may need separate review or stay out, as the guide to solar installers' project records notes, and a manufacturer's decisions on a customer's claim may sit under confidentiality terms. Removing names and site addresses lowers privacy risk but does not create a right to license a record the company does not control. Expect each licensor to support some labels well and others barely: an installer's file can teach Request evidence and Approve with conditions convincingly, while Escalate cases usually need the warrantor's engineering notes.

How to scope a request for warranty completeness cases

Ask for decided claims with their evidence inventories, not for warranty data in general. A request that names the decision, the five labels, the field groups above, a failure-window coverage measure and an acceptance test lets a supplier say quickly whether its records can carry the task.

At a glance

How to scope a request for warranty completeness cases

SourceX licenses records from the companies that hold them and supplies them to AI developers in its own name, sourcing each dataset when a buyer asks rather than keeping stock. Its request form asks for the domain, the source systems, licensing requirements such as exclusivity, duration and permitted use, and de-identification requirements, which is where a warranty team's intended use and acceptance test belong. Whether a given installer, operations contractor or manufacturer can supply these cases is settled in review, one company at a time. What to put in the request:

Checklist · 5 items

  1. Decision and population: commercial inverter and module claims, the date range, and whether claims still open are included.
  2. Label coverage: a minimum count per label, with hard negatives in both directions.
  3. Evidence fields: the evidence inventory with holders, the terms version and failure-window coverage.
  4. Outcome fields: final decision, failure analysis result and recurrence, with no-response closures kept.
  5. Acceptance test: the rubric above run on a held-out sample, with reviewer agreement reported.

0 of 5 ticked. Tick each item as you go.

Methods note: the PVDAQ count, the standards fields and the casebook

Data: systems_20250729.csv from the PVDAQ folder of NREL's Open Energy Data Initiative data lake (S3 bucket oedi-data-lake, last modified 31 July 2025), and a listing of every object under pvdaq/csv/pvdata/system_id=<id>/ for each system, taken through the public S3 listing endpoint on 28 October 2026. Three standard-library scripts written for this article, pvdaq_listing.py, pvdaq_count.py and pvdaq_strips.py (the three systems in the figure), fetch or read only those files, use no randomness and run no repository code; they are kept with their output in the writer's working folder, with a summary table that holds system ids and computed fields only.

Rules as applied: the systems file is read as UTF-8 with any byte-order mark stripped (it has none); a system is a distinct system_id after trimming spaces, keeping the first row for the two ids that appear twice. Span is last_timestamp minus first_timestamp, parsed as month/day/year hour:minute; an impossible, blank or reversed pair leaves the span unknown, which happens for two systems with blank timestamps. A short-history note is a qa_issue cell containing the text less than 1.0 years data, matched without regard to case, which covers both wordings; two years means 730 days or more. A daily-file system has at least one non-empty CSV file at a year, month and day path and no other non-empty file. Each such file gives a calendar date, and dates that do not exist are rejected (none occurred). Dates earlier than the calendar date of the system's first timestamp are dropped as clock faults: eight dates on two systems. The record runs from the first to the last remaining date; a gap is a run of consecutive record days without a file; coverage is days with a file divided by record days. A file's presence does not show that the day is complete, and file sizes were not used. The median is exact and percentages are rounded once.

The standards fields come from Master-OB-OpenAPI.json in the Orange Button taxonomy at commit f20d324 and json/model_103.json in sunspec/models at commit 90b4a33, counted by standards_fields.py as the keys of each object's properties; a schema mentions a claim when the word appears anywhere in its definition. The casebook, field groups, routing rule and rubric are editorial frameworks drawn up for this page on 28 October 2026; the cases are invented, and xs0822_casebook.py applies the routing rule to them. Not measured: how often real claims fall under each label, the cost of over- or under-asking, reviewer agreement, or whether models trained on such cases decide better. PVDAQ was read from NREL's own data lake, which is refreshed, so a rerun may differ.

Frequently asked questions

What is a warranty evidence completeness dataset?

It is a set of decided warranty claims in which each case records which evidence was present, missing or conflicting, whether each gap could change the answer, and the label the desk chose. It teaches a model to ask for the one record that matters and to decide without it when no record could change the outcome.

Does a long monitoring history prove a solar warranty claim?

No. What matters is whether the failure window, and any period a performance claim depends on, was actually observed. In NREL's PVDAQ archive, 66 of the 131 daily-file records spanning two years or more contain a gap of 30 days or longer.

Why include claims that were approved despite missing records?

Because a model has to learn that some gaps are harmless. Without approved cases that had irrelevant gaps, it learns to request documents whenever anything is blank, which delays valid claims and raises handling cost.

Can public PV monitoring data train a warranty model?

Public archives such as PVDAQ are useful for testing coverage logic and failure-window checks, but they hold no claims, terms or decisions. Decision labels have to come from warranty desks, installers or operations contractors whose records include outcomes.

Who should label warranty completeness cases?

A warranty or applications engineer and a contracts reviewer, labeling independently, with an adjudicator for disagreements. Keep agreement per case: low-agreement cases are valuable for evaluation but should not be treated as settled training labels.

Sources

  • NREL's pvdaq_access repository says data from public PV sites is available through the OEDI data lake, refreshed monthly where possible and stored in daily CSV or Parquet files per system. Source
  • PVDAQ's systems file dated 29 July 2025, counted for this article: 1,860 distinct systems, first and last timestamps, and a quality-assessment column with less than 1.0 years of data notes. Source
  • The PVDAQ CSV data folder in NREL's OEDI data lake, listed for every system on 28 October 2026 to measure gaps in daily-file monitoring records. Source
  • PVDAQ's July 2025 change log records 1,454 systems added from the PV Output group. Source
  • The Orange Button solar taxonomy defines Device, Warranty and FailureRecord objects and no object for a warranty claim or its decision. Source
  • SunSpec's three-phase inverter model defines a mandatory 32-bit event field with 16 flags, such as GROUND_FAULT, AC_OVER_VOLT and HW_TEST_FAILURE. Source
  • NREL's 2015 SAPC best-practice guide asks owners to keep the plant history, warranties and an equipment serial number inventory, and to retain original-resolution monitoring data for the life of the system. Source
  • NREL's third-edition PV operations and maintenance best-practices guide treats curating equipment inventories, manuals and warranties as part of administering operations. Source
  • IEC 62446-1 defines the documentation handed over after a grid-connected PV system is installed, with commissioning tests and inspection criteria. Source
  • IEC 61724-1:2021 covers terminology, equipment and methods for PV performance monitoring and defines classes of monitoring systems. Source
  • UCC section 2-607 requires a buyer that accepted goods to notify the seller of a breach within a reasonable time or be barred from any remedy. Source
  • UCC section 2-725 says a warranty breach occurs at tender of delivery unless the warranty explicitly extends to future performance. Source
  • SourceX's request form asks buyers for domain, source systems, licensing requirements and de-identification requirements, and says datasets are sourced on request, not held in stock. Source

Tell us what data you need

Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.

See if your company qualifies