Methodology

How accuracy and confidence scores are produced

  1. 01

    Holdout testing

    Model outputs are checked against a set of photos the models were not trained on, so reported accuracy reflects performance on unseen listings rather than photos the system has already seen.

  2. 02

    Human-reviewer calibration

    A sample of outputs is periodically checked against a human reviewer's independent judgment on the same photos. Disagreements are used to recalibrate confidence thresholds, not just logged and ignored.

  3. 03

    Confidence-based routing

    Outputs below a set confidence threshold are routed for a quick human check rather than delivered as-is. The threshold itself is tuned against the calibration results above, not fixed arbitrarily.

  4. 04

    Ongoing re-testing

    As the photo set grows, the holdout test is refreshed periodically so reported accuracy reflects current performance rather than a one-time benchmark.

NOTE This describes the validation approach the platform is built around. Before this page or any accuracy figures are used in an external application or pitch, confirm the current holdout size, calibration cadence, and reviewer process against what's actually running today.

Known limitations

Where a human should look twice

Concealed conditions

Anything behind walls, above ceilings, or below grade (plumbing runs, wiring, foundation issues) isn't visible in a photo and isn't something the system can assess.

Poor or inconsistent photos

Dark, blurry, or heavily staged/edited photos reduce confidence across the board. Low-confidence outputs are flagged for review rather than guessed at.

Unusual or non-standard layouts

Rooms with atypical layouts or mixed use (a bedroom doubling as an office, for instance) are harder to classify cleanly and warrant a second look.

Regional and style variation

Condition and quality norms vary by market and property style. A score should be read as a standardized signal, not a substitute for local expertise.

Data handling

What happens to the photos you upload

Ownership

Photos uploaded for analysis remain the property of the brokerage that uploaded them. Stryder processes them to produce the structured output described on the Technology page.

Retention

Photos and their derived analysis output are retained for as long as your account is active, so past listings stay accessible. If an account is closed, data is deleted within 90 days, except where retention is required to meet a legal obligation. Earlier deletion of a specific listing's data can be requested at any time. See the full Privacy Policy for details.

Bring a listing you already know the answer for

The fastest way to judge accuracy is to run it against a property your team has already reviewed by hand, and compare the two.