UAD 3.6 appraisal QC: AI models that learn your own review rules

The 2 November 2026 mandate introduces UAD 3.6 and the redesigned URAR for new appraisal submissions to the Uniform Collateral Data Portal. A temporary policy exception allows lenders that request it to continue submitting UAD 2.6 reports through May 19, 2027. See Fannie Mae’s current transition guidance. The ratings scales are the same, but almost everything around them is more structured: photos belong to specific sections, condition and quality are rated separately inside and out, and defects are recorded as data where they occur.

More structure means more to check on every report. Lenders' review platforms already run automated checks against the GSE standards. Your own review history can add another layer: the revisions your reviewers send back, the photos your QC team rejects, where your chief appraiser draws the line between C3 and C4.

That's what this post is about: turning the decisions your team has already made into AI models that make the same calls on the next report, and send the unsure ones to a person.

For a reviewer, that's more photos to match to sections, two sets of ratings to check against them, and defects that have to appear in the right place.

Each of these is a decision your team already makes many times a week. That's what makes it a good model: there are past answers to learn from.

1. Revision requests: what kind, and who fixes it

When a reviewer writes a revision request, someone has to decide what kind it is and where it goes. A text model learns that from your past requests:

One model can answer both parts at once, the category and who it goes to, because StayCharted models can give an answer with several parts. Your revision log becomes a report: which categories come up most, for which appraisers, and how often.

2. Photo checks: retake or accept

Your QC team already sends photos back: too dark, blurry, people or personal items in shot, the wrong room. A picture model learns your accept-or-retake line from photos the team accepted and rejected:

• Accept
• Retake: too dark or blurred
• Retake: people or personal items visible
• Retake: doesn't show the area

3. Photo sections: is this photo where it says it is?

Train a picture model on photos from your past reports, with the section each one was filed under as its category: front, rear, street scene, kitchen, bathroom, bedroom, living area, below-grade, outbuilding, data plate. On a new report, the model labels each photo, and anything it confidently places in a different section from the appraiser's is worth a look. So is a report with no photo the model calls "bathroom" when the property has two.

The comparison is a simple rule in your QC checklist or system; the model supplies the label and how sure it is.

4. Condition grades your way

Apply the GSE definitions consistently; house review rules must not redefine them. Have reviewers label the condition evidence visible in each photo. Do not automatically give every photo the report’s overall rating: a room photo may not support that judgment. A model trained on those reviewed examples can flag photos for a second look.

Use it as a second opinion, per photo: a report rated interior C2 where most of the kitchen and bathroom photos read C4 is one to look at first. The appraiser's rating is the rating.

For defects, damages and deficiencies, keep it simple: one yes-or-no model per item your reviewers care about, such as water staining, roof damage or a missing handrail, trained on photos where your team flagged it and photos where it didn't.

• Past decisions, labelled the way your team made them. Your report archive already has most of it: every photo filed under a section, every report with its ratings, and your revision log with what each request was about.
• For an initial pilot, 20 to 50 examples per category to start, and more for categories that look alike, such as C3 and C4.
• Enough of the rare ones. C5, C6 and serious defects are uncommon, so collect them on purpose. A model can't learn a category it has barely seen.

1. Pick one check. Revision categories are the easiest start: text only, and your log already has the answers.
2. Upload your examples. For text, a spreadsheet with the request and its category (and who handled it, for a two-part answer). For photos, a ZIP with one folder per category, or a spreadsheet with a link to each photo and its category. A picture model can also read a note alongside each photo, such as the section it was filed under.
3. Let StayCharted check the data. Before training it flags rows that look like they belong in another category, the same photo under two categories, near-identical shots and files it couldn't read.
4. Train, then read the report per category. Start with the AI Classifier Model for text and the AI Image Classifier Model for photos: about a minute each, no GPU. Look at accuracy per category, not only overall. A model can be right 95% of the time overall and still miss most of the C5s.
5. Connect it. Call the model through the API from your QC system (up to 20 photos a call, each with an answer, a confidence and a review flag), fill a spreadsheet of a week's reports at once, or ask it from Claude or ChatGPT through StayCharted's connector.
6. Set a review cutoff. On Essentials and Business, answers below it wait in the Review Queue for a reviewer. Their corrections become training examples for the next version.

• The same property in training and testing. Ten photos of one house split across both makes the test look better than real life. Keep each report's photos together.
• Giving the model the answer. Don't include the appraiser's name, the rating or the property address as a column it learns from. It will learn "this appraiser usually rates C3" instead of reading the photo.
• Mixing questions in one model. "Kitchen" and "C4" are different questions. One model per check.
• Training on unresolved disagreements. If two reviewers would label the same photo differently, settle the rule first. A model can't be more consistent than the rule it learns.
• Never retesting. Retest after your guidelines change. UAD 3.6's rewritten rating definitions are a good reason to retest now.

• Unsure answers go to a reviewer, and confident answers can still be wrong. Sample confident answers too. You choose the review cutoff.
• Ratings stay the appraiser's. The model gives a second opinion on a photo; it doesn't change the report.
• Revision requests are still written by people. The model files them and routes them.

• It doesn't produce a UAD 3.6 report or XML. Your form software does that.
• It doesn't measure. It won't give square footage or room dimensions.
• It doesn't check the GSE rules for you. It learns your team's decisions; the GSE requirements are your checklist and your review platform's job.
• It doesn't know categories it hasn't seen. A new revision type needs examples first.
• It isn't a replacement for the appraiser or the reviewer. It does the first pass.

• Fannie Mae: UAD transition and policy exception
• Freddie Mac: UAD 3.6 FAQs and rating changes
• GSE Photo and Image Requirements job aid

Photo requirements depend on report applicability. Use the official job aid for required, conditional and optional photos.