Systems and records
Property maintenance evidence dataset: when tenant reports disagree
By SourceX Editorial · Reviewed by Noah Loul ·
9 sections
Short answer
A property maintenance evidence dataset should store a disputed repair report as competing claims, each with its source, time and basis, and label every decision point with one of seven actions: make safe now, escalate, re-scope, inspect first, request evidence, close with a record, or dispatch as reported. Hackney Council's public repairs code shows the gap: its closing form records only visit completed or no access.
System export
Property maintenance
Key takeaways
- Label the next action at each decision point, not a verdict on who was right: seven actions, tested in an order that puts safety first and closure late.
- Store every account as its own claim with source role, basis and time; an inferred cause such as lifestyle must never outrank a reading.
- Hackney's public Repairs Hub code records no access and scope changes explicitly, but its two closing reasons cannot separate a fix from a visit that found nothing.
- Score decisions blind to outcome, keep disagreements as contested labels, and treat runs of failed access as unknown rather than resolved.
- Health and vulnerability details are special category data under UK law; drop them or keep a coded flag with a recorded lawful basis.
What decision does a disputed repair report force?
It forces the next action, not a verdict on who is right. When one tenant says water is coming through the ceiling and the neighbor above says nothing is leaking, or an operative closes a heating job that the tenant reports again two days later, the manager still has to choose what happens next and how fast. The useful label is that choice, the evidence it rested on and the deadline it ran against.
Deadlines make the choice concrete. In England, Awaab's Law already requires social landlords to act on damp and mold and on emergency hazards within set times: GOV.UK's guidance gives 10 working days to investigate a significant hazard and 5 working days after the investigation to start the safety work, while emergency hazards must be investigated and made safe within 24 hours of the landlord becoming aware of them. Its second-phase guidance extends the duties from 30 November 2026 to excess cold and heat, falls, structural collapse, fire and explosions, electrical hazards and domestic hygiene. Because the emergency clock runs from the moment the landlord becomes aware, a dataset has to record when each account arrived, not only what it said.
Outside English social housing the deadlines differ, but the problem has the same shape: two or more accounts of one condition, an action that has to be taken before they agree, and a record that usually keeps only one of them. This page sets out the actions, counts what one council's public repairs code records when a report meets the site, and builds a claim-level packet, a casebook and a reviewer's checklist.
Seven actions, and the rule that picks one
Test seven actions in a fixed order and take the first that fits: make safe now, escalate, re-scope, inspect first, request evidence, close with a record, and dispatch as reported. Safety comes first because a conflicting account is never a reason to lower a priority; closing comes late because it needs the most evidence.
Two boundaries do most of the work. Inspect first applies when the accounts disagree about the cause or severity of a possible hazard, no resident's photo, reading or access date would settle it, and only a diagnosis on site can, as with recurring mold that an old visit note blames on condensation. Request evidence applies when no uncontained hazard remains and a photo, a reading or an access date from a resident would settle the conflict, with a date set to decide again. Closing needs a completed inspection that found no defect and findings the reporter has received; a theory about the cause is never enough.
Each action, and what the record has to hold when it is taken:
| Action | Use when | The record must hold | What it prevents |
|---|---|---|---|
| Make safe now | Any credible report describes an emergency hazard, whatever the other accounts say | The report, the time it arrived, the containment step | Waiting for agreement while a hazard stays live |
| Escalate | The cause or the fix sits outside the landlord's own trades or authority | The other owner, engineer, insurer or service, and the referral time | Ordering work the landlord cannot do or authorize |
| Re-scope | Site evidence shows a different cause or extent than the report or the order | The operative's finding, the changed job lines, the approval | Repairing the symptom and leaving the cause |
| Inspect first | Accounts conflict on the cause or severity of a possible hazard, and no resident's photo, reading or access date would settle it | Both accounts, earlier visits, a surveyor booking | Closing recurring damp on a lifestyle theory |
| Request evidence | No uncontained hazard remains, and a resident's photo, reading or access date would settle it | What was asked for, from whom, and a re-decision date | An open-ended wait that nobody owns |
| Close with a record | A completed inspection found no defect and the reporter has the findings | The inspection result, the notice sent, the route to report again | A silent closure that returns as a complaint |
| Dispatch as reported | Report and records agree, or the conflict changes neither the trade nor the priority | The report and the records that match it | Needless inspections on routine jobs |
What a council repairs system records when a report meets the site
Hackney Council's Repairs Hub, whose code the council publishes on GitHub, records failed access and changes of scope explicitly but has no closing value that separates a repair from a visit that found nothing wrong. The frontend repository's README says agents and managers use it to raise, review, approve and close work orders, and that operatives complete their jobs in a mobile view. A script read its source files at commit 9612b7a (27 October 2026), the latest of 2,499 commits since 26 November 2020, without running any of its code.
The code defines eight work order statuses, five of which record a job that stalled or changed: no access, authorisation pending approval, and a variation that is pending, approved or rejected. The operative's closing form offers two reasons, visit completed or no access, which the code sends as HACT job status codes 0 and 70. A completed visit can then record whether further work is required and, if so, in which of 13 trades; two of them are a surveyor's inspection and a leaks hub. A change of scope travels as a variation, sent as HACT code 80, more specific schedule-of-rates codes identified, which a supervisor or the repairs team approves or rejects.
The raising side is thinner. The repair description an agent types is capped at 230 characters. Before commit 4732a7c on 27 October 2026, the reason-for-repair question offered damp and mold or none of the above; that commit adds a ten-reason list, shown only when a feature toggle is on, matching damp and mold plus the second-phase hazards in GOV.UK's guidance, with falls split three ways. So the system can say that a visit happened, that nobody got in or that the job changed, but not whether the tenant's account was confirmed, and a second account of the same condition has no structured place to go.
The decision points, and the choices the code offers at each:
| Decision point | Choices in the code | Count | What the record cannot say |
|---|---|---|---|
| Raising: reason for repair | Damp and mold, or none of the above; a ten-hazard list behind a feature toggle | 2, or 11 with the toggle | Whether the hazard was confirmed or only reported |
| Raising: repair description | Free text typed by an agent | 230 characters | A second account of the same condition |
| Priority | Immediate, emergency, urgent, normal, survey, Awaab's and six other codes | 12 codes, 4 needing appointments | Why that priority was chosen |
| Closing a visit | Visit completed or no access | 2 | Whether the fault was fixed or never found |
| Further work | None, or further work in one of 13 trades, including a surveyor's inspection | 2 answers, 13 trades | Whether the follow-on contradicts the original report |
| Change of scope | Variation pending, approved or rejected, sent as HACT code 80 | 3 statuses | Who first suspected a different cause |
| Photos | Close without a photo for one of four reasons | 4 | What a photo would have shown |
Keep the conflict: one row per claim, not one description
A dataset that resolves the dispute before storing it teaches a model to pick a side. Keep each statement about the condition as its own claim, with who made it, what it rested on and when, and attach every decision to the claims that existed at that moment. The basis field matters most: a measurement or a direct observation should never be outranked by an inferred cause.
That rule has an official precedent. The Housing Ombudsman's October 2021 spotlight report on damp and mold, titled It's not lifestyle, made 26 recommendations, urged landlords to stop blaming residents' lifestyles for damp and called for a shift from reactive to proactive handling. A record that stores a lifestyle note as the cause, with no basis attached, carries exactly that habit into any model trained on it.
The fields, why the label needs each one, and how each is treated before records leave the landlord:
| Field | Content | Why the label needs it | Treatment before release |
|---|---|---|---|
| Claim | One row per statement about the condition, verbatim where possible | The action rests on which claims conflict | Names, third parties and health details redacted from the text |
| Source role | Tenant in the unit, neighbor, operative, surveyor, sensor, staff | Conflicts run between roles | Role only, never a name |
| Basis | Seen, measured, reported second-hand or inferred | An inferred cause must not outrank a reading | Kept as is |
| Times | When the condition was seen and when it was logged | Deadlines run from awareness | Dates kept; clock times generalized if they identify anyone |
| Location | Unit, room, building element, and the units above or beside | Leaks and damp cross unit lines | Units coded; addresses removed |
| Decision point | The action, the role that took it, the time and the deadline that applied | Each decision point is one example | Staff names replaced by roles |
| Outcome evidence | Follow-on job, variation, inspection result, recurrence, complaint result | The outcome label comes from here | Documents redacted; photos screened for faces and papers |
| Vulnerability flag | Whether a household vulnerability changed the priority | A different priority can be the right one | Health data is special category data; keep a coded flag only with a lawful basis, else drop it |
How is the right action known? Outcome labels from later records
Score each decision against what the evidence on file supported at the time, and store what happened next as a separate outcome. Two people with repairs experience, such as a surveyor and a repairs manager, should label each decision point without seeing the outcome. When they disagree, keep both labels and their reasons and mark the point contested instead of forcing a vote; contested points make good evaluation items, because they test whether a model asks for the evidence that would settle them.
At a glance
How is the right action known? Outcome labels from later records
Later records supply the outcome, and each source has its own trap:
- Recurrence: a new report on the same element within a set window says the action did not hold; the window is a choice to state, not a standard.
- Follow-on or variation: the first order had the wrong scope, and the record shows which way it was wrong.
- Inspection findings: moisture readings, photos or a surveyor's report settle cause and severity when they exist.
- No access: a run of failed visits means the condition was never verified, so the outcome is unknown rather than resolved.
- Complaint or ombudsman outcome: an outside judgment on how the report was handled, arriving months later.
- Confirmation: the reporter agrees the problem has gone, which supports a closure but can be missing for reasons that have nothing to do with the repair.
Eight disputed reports and the action each deserves
The casebook applies the fixed order to eight invented reports. A short script assigns each action from the listed facts, so every verdict comes from the seven actions and no others.
Case K1 also shows why each decision point, not each case, is one example. On day 0 the manager makes safe. On day 1, with the water contained, the open question is whether the leak starts in flat 4B, and an access date from that tenant would settle it: request evidence. On day 4 an operative gets in and finds a failed bath waste in 4B, so the order is re-scoped from a ceiling repair in 3B to a pipe repair upstairs. Three decisions, three labels, each scored against the evidence on file that day.
| Report (illustrative) | Conflicting evidence | Action | Deciding fact |
|---|---|---|---|
| K1: water dripping near a ceiling light in flat 3B | The tenant in 4B says nothing is leaking; the first visit cannot get into 4B | Make safe now | Water near electrics, whatever 4B says |
| K2: black mold back in a bedroom after a wipe-down | An earlier visit note blames drying clothes indoors | Inspect first | The cause is disputed and only a diagnosis settles it |
| K3: one radiator cold, the rest of the flat warm | The thermostat reads normal in the hallway | Dispatch as reported | The reading is from another room; the trade does not change |
| K4: blocked kitchen sink | The operative finds a cracked waste pipe inside the wall and a damaged cabinet | Re-scope | Site evidence shows a different cause and extent |
| K5: damp patch spreading on a ceiling | The flat above belongs to a leaseholder, not the landlord | Escalate | The source sits in property the landlord does not control |
| K6: window will not close, reported in an app with no photo | Nothing on file either way; no hazard described | Request evidence | A photo and an access date would settle it |
| K7: damp patch reported in a hallway | A surveyor's readings found an old, dry stain; the findings were sent | Close with a record | A completed inspection found no defect |
| K8: a neighbor smells gas outside flat 2A | The tenant in 2A, reached by phone, smells nothing | Make safe now | A credible gas report outranks a denial |
Hard negatives that pull a triage model the wrong way
An evaluation set needs cases where the obvious reading of the record is wrong. Each item below punishes a model that trusts a field, a word or a status instead of the evidence behind it:
- A thermostat reading that contradicts a heating complaint but was taken in a different room.
- A visit note that names condensation or lifestyle as the cause with no reading behind it; it is a theory, not evidence.
- A neighbor's denial of a leak after a visit that never got in: no access is not no leak.
- Twelve reports from one block in an hour after a burst main: one event and one decision, not twelve conflicts.
- A report that says flooding with a photo of a dripping trap: a normal-priority repair, not an emergency.
- A tenant who reports often: frequency is not a reason to lower a priority, and a model that learns otherwise has learned a bias.
- A completed status with no further work required after a visit that found nothing: the record looks fixed, but nothing was repaired.
A domain review checklist before a case enters the set
Before a case enters a training or evaluation set, a reviewer with housing repairs experience should confirm every item below. The same list works as the acceptance test a buyer attaches to a data request:
Checklist · 8 items
- Every account is stored as a separate claim with its source role, basis and both times.
- Each decision point carries only the claims that existed when it was taken.
- The deadline that applied, and the event that started it, are recorded for every hazard case.
- Labels were assigned by two reviewers who had not seen the outcome; disagreements are kept and marked contested.
- Outcome fields come from follow-ons, variations, inspections, recurrence and complaints, never from the status alone.
- No-access runs are labeled unknown, not resolved.
- Names, addresses, contact details and identifying photo content are removed; health and vulnerability details are dropped or reduced to a coded flag with a recorded lawful basis.
- The intended use is written down, and it excludes deciding repair liability or a tenant's entitlement without an accountable person.
0 of 8 ticked. Tick each item as you go.
What a landlord's records can and cannot show
A landlord's or property manager's records can show which accounts arrived and when, what was ordered, what operatives found and how disputes ended. They cannot show conditions nobody reported, and the Ombudsman's spotlight report warned that many landlords rely too heavily on residents to report problems, so a dataset built from reports inherits that gap. Nor can they show why a call handler chose a priority unless someone wrote it down.
A sector label settles none of the rights questions. Tenancy terms, privacy notices, contractor agreements and the terms of the housing management software each need review, and tenant records hold personal data and sometimes health data, which the ICO treats as special category data, including health information inferred from other data. Keep the intended use bounded: triage support and evidence requests, tested on contested cases, rather than automated decisions about who pays.
The Repairs Hub figures come from a short script that read the frontend's source files as text at commit 9612b7a on 28 October 2026 without running them. Statuses are the exported STATUS constants with a code and a description; option counts are the text labels inside each named option list; priorities are the exported integer constants in the priorities helper, and the four needing appointments are the names in its list of codes requiring appointments; HACT codes are the quoted numbers on lines that assign a type code; the 230-character limit is the maxLength set on the description field, and the toggle finding is the conditional that renders the ten-reason list only when the toggle is on. The commit count is git's count of commits reachable from 9612b7a, and the date of the ten-reason list comes from that file's history with renames followed. Not measured: how often any status, reason or priority occurs in Hackney's real work orders, which the frontend does not hold, or what the backend stores beyond what the frontend sends.
SourceX sources property maintenance records when a buyer asks for them rather than holding them in stock. A useful request names the decision points to label, the action set (these seven or your own), the packet fields, the jurisdictions and years needed, and an acceptance test, such as agreement with two adjudicators on a held-out casebook of contested reports.
Frequently asked questions
What is a property maintenance evidence dataset?
It is a set of repair cases in which every account of a condition, from tenants, neighbors, operatives, surveyors or sensors, is stored as a separate claim with its source, basis and time, and each decision point is labeled with the action taken. The outcome from later records is kept in its own field, so a lucky decision does not become the label.
Why not label each repair case with a single category?
A disputed report produces several decisions as evidence arrives: make safe, ask for access, re-scope. One category per case hides that sequence and mixes evidence the manager did not yet have into the label. Label each decision point with the claims that existed at that moment.
Can a model decide who is right when tenants disagree?
It should not be asked to. The useful task is choosing the next action and naming the evidence that would settle the conflict, such as a reading, a photo or an access date. Liability and entitlement decisions need an accountable person and the full facts.
Does Awaab's Law apply to private landlords?
As of this article's date, the duties apply to social housing landlords in England, and GOV.UK has published a separate consultation on extending Awaab's Law to private rented tenancies and licenses. Check the current guidance and take legal advice before relying on any deadline in a dataset rule.
What should be removed from tenant repair records before licensing?
Remove names, addresses, phone numbers, staff names and identifying content in photos, and redact third parties from free text. Health and vulnerability details are special category data under UK data protection law, so drop them or keep only a coded flag where a lawful basis is recorded.
Sources
- Hackney Council's public Repairs Hub frontend at commit 9612b7a (27 October 2026): the README says agents and managers use it to raise, review, approve and close work orders and operatives complete jobs in a mobile view; the history holds 2,499 commits from 26 November 2020. Source
- Repairs Hub status codes: eight work order statuses including No Access, Variation Pending Approval, Variation Rejected and Authorisation Pending Approval; two closing reasons (visit completed, no access); two further-work answers; 13 follow-on trades including Surveyors Inspection and Leaks Hub. Source
- Repairs Hub priority constants: twelve codes including immediate, emergency, urgent, normal, survey and Awaab's, with urgent, normal, survey and Awaab's requiring appointments. Source
- Repairs Hub raise form: a repair description limited to 230 characters; a reason-for-repair question offering damp and mold or none of the above, and a ten-reason list added on 27 October 2026 that shows only when a feature toggle is on. Source
- Repairs Hub HACT builders: closing a work order sends job status code 0, or 70 for no access; a variation sends code 80, more specific SOR codes identified. Source
- GOV.UK guidance for social landlords on Awaab's Law timeframes: significant hazards investigated within 10 working days and safety work within 5 working days of the investigation; the law currently covers damp and mold and emergency hazards. Source
- GOV.UK Awaab's Law Phase 2 guidance: emergency hazards investigated and made safe within 24 hours of the landlord becoming aware; from 30 November 2026 the duties extend to excess cold, excess heat, falls, structural collapse, fire and explosions, electrical hazards and domestic hygiene. Source
- Housing Ombudsman spotlight report on damp and mold, It's not lifestyle (October 2021): 26 recommendations, a call to stop blaming residents' lifestyles and to move from reactive to proactive handling, and a warning that many landlords rely too heavily on residents to report. Source
- GOV.UK consultation page: Extending Awaab's Law to private rented tenancies and to licences. Source
- ICO guidance: health data is special category data under the UK GDPR and covers information that reveals or concerns a person's past, current or future health, including by inference. Source
Tell us what data you need
Share scope, volume, format, timing and licensing needs. Datasets are sourced on request.