METHODOLOGY.md

# Labelling Methodology

This document explains how the labels attached to every image in this sample are produced, who produces them, and what quality guarantees do and do not apply. It is written for technical evaluators; nothing here is marketing.

## 1. Who applies the labels

Every image is captured during a professional property inspection (inventory, check-in, or check-out) of a UK residential rental property, carried out by a working inventory professional on behalf of a letting agency or property manager. Labels are applied by the inspector at the point of capture, in the room, on the device — they are not retrospective annotations by third-party labellers.

Inspector identity is recorded against every photograph and every report in the platform. Individual inspector credentials and training vary by the agency they work for and are not held in the miProgram database.

The `annotator.role` field in each image's metadata reflects the platform access role assigned to the user by their agency (for example, "Client Admin", "Client User") — these roles govern permissions within the platform, such as edit and delete rights, and are used to prevent data loss when personnel leave an organisation. They do not describe the user's professional job. All photographs in this sample are captured during professional property inspections carried out on behalf of a client agency by working inventory professionals.

## 2. Structured, picklist-driven capture

The mobile app that captures inspections is dropdown-driven. Condition ratings, room types, materials, colours, finishes, defect types, brands, and every other label in this sample come from centrally-maintained reference lists (see `TAXONOMY.md` for the full enumeration of all 42 lists). Inspectors select values; they do not type them. Free-text exists in the platform only as supplementary notes fields, none of which are included in this sample.

Practical consequences for training data:

- Label vocabulary is closed and consistent across the entire dataset — no spelling variants, no synonyms, no free-text noise.
- Label completeness is structurally enforced: in the February 2026 production window this sample is drawn from, condition, material/type, and colour fields were populated on effectively 100% of captured elements.
- Labels reflect the judgement of the professional on site. Condition ratings are one person's assessment made for a contractual inventory document, not a consensus of multiple annotators. No inter-rater agreement measurements are available.

## 3. What a photograph is attached to

Photographs are captured at specific levels of the inspection hierarchy, and their metadata reflects that level:

- **Room overview photographs** are attached to a room. They carry the room type and property context, but no condition rating — condition is recorded per element, not per room. In these images `labels.condition_rating` is `null` by design.
- **Element photographs** (walls, floors, ceilings, doors, windows) are attached to a specific element that carries a condition rating, material/type, colour, finish where applicable, and zero or more defect labels. These are the fully-labelled records.
- **Product-level photographs** (appliances, fixtures) are attached to an item record that can carry a brand, type, colour, and condition.

## 3a. Implicit default conditions

Not every item category in the platform carries an explicit condition field. The five element categories in this sample's condition and defect strata — walls, floors, ceilings, doors, and windows — always carry an explicit, inspector-selected condition rating, and product-level items (appliances, typed fixtures) carry one as well.

For certain other item categories — furnishings, crockery, glassware, utensils, and count-based fixture items such as plug sockets, light switches, and sinks — the platform does not capture a per-item condition rating at all. For these categories, the client-facing inventory document renders the condition as "Good" by default; the inspector records departures from good condition through defect labels and notes rather than a rating field. This is a deliberate workflow optimisation that reduces on-site data entry time for items that are visibly in good condition.

Consequently, in the full dataset, images of items in those categories carry no `condition_rating` value at the database level, and should be interpreted as **implicitly Good** unless accompanied by defect labels or notes indicating otherwise. The true production frequency of "Good" condition is therefore higher than explicit-Good counts suggest, and for downstream use a `null` condition on these categories may be treated as a weakly-supervised Good label. (In this sample, `condition_rating: null` appears only on room-overview photographs — where no rating exists at any level — and this default does not affect the condition-rating stratum, which contains only explicitly-rated elements.)

## 4. Quality assurance

QA in the platform operates at **report level, not per image**. A completed inventory is compiled, optionally reviewed under an approval workflow, and sent to the client agency, at which point it becomes the contractual record of the tenancy and is locked. The `annotator.report_qa_status` field in each image's metadata reflects the parent report's stage in that workflow at the time this snapshot was taken ("Draft", "Awaiting approval", "Sent to client"). There is no per-photograph verification flag, and this sample does not claim one.

Reports are also subject to challenge by landlords and tenants during and after the tenancy (the platform records tenant disagreement against individual elements), which creates a commercial feedback pressure toward accurate labelling: these documents are used to adjudicate real deposit disputes.

## 5. Image provenance and processing

- **Source**: real production inspections of vacant or tenanted UK rental properties, captured on inspectors' own iOS devices (the EXIF device field is preserved and shows the real device mix, iPhone 8 through iPhone 17 generations plus iPads).
- **Resolution**: the platform's sync pipeline transmits images at reduced resolution; the dominant stored format is 480x640. A small fraction (roughly 1–2%) of images are stored at native camera resolution (3–5 megapixels). This sample is drawn representatively and reflects that profile honestly — see the README's composition notes.
- **Re-encoding**: every stored image has been through one JPEG re-encode at quality 95 during upload processing (orientation is normalised at the same step). These are second-generation JPEGs, not raw camera output. This is standard practice for mobile-captured production imagery at scale.
- **EXIF**: capture device model and original capture timestamp are preserved where the device recorded them. `capture_date` in the metadata comes from EXIF and occasionally reflects an incorrect device clock or a camera-roll import; `upload_date` is the platform-recorded date and is authoritative. **GPS coordinates and all location EXIF were checked for and stripped from every image during sample preparation.** File names and metadata contain no property identifiers; geography is expressed only as an ITL1-style region derived from (and replacing) the postcode.

## 6. Privacy and exclusions

- Every image in this sample was individually reviewed for personally identifying content. Images containing people or their reflections, readable post or documents, personal photographs, street signage identifying the specific street, or vehicle registration plates were excluded, regardless of image quality. House numbers may appear on exterior property elevations; because all inspections in this sample were conducted on vacant properties, these identify a property rather than any individual and are not personal data under UK GDPR.
- **Mid-term (occupied-tenancy) inspection photographs are structurally excluded.** In the platform schema, mid-term inspections write to physically separate tables from inventories, check-ins, and check-outs; this sample is drawn exclusively from the non-mid-term tables, so the exclusion is architectural rather than a filter.
- One room type ships empty in this sample for composition rather than data availability: open-porch photographs from the source window were predominantly close-range doorway shots dominated by the house number in frame or otherwise off-subject; none met the selection bar under the conservative v1 review. Larger open-porch sets are available for commercial supply.

## 7. Known limitations, stated plainly

- One annotator per label; no inter-rater agreement statistics.
- Condition-rating usage is heavily skewed toward "Good" in production (as it is in reality); rare ratings are represented with fewer exemplars, with counts stated in the README.
- No per-image QA flag exists.
- `capture_date` inherits device-clock quirks; `upload_date` is authoritative.
- Images are second-generation JPEGs at predominantly 480x640.

Everything in this section is also true of the full dataset. The sample is drawn to be representative of it, not to flatter it.