Verified extraction means every score that appears in an AI-drafted report has been matched back to the document it came from, and the descriptive label beside it comes from a fixed table rather than from the AI's own wording. In plain terms: the number on the page is the number in your file. It doesn't mean the file itself was right, and it doesn't replace your read before you sign.
I want to explain why we built it, because the reason is less obvious than it sounds. Most people assume software is good with numbers. The kind of AI that writes prose is, in an important way, not.
I'm not a clinician, and nothing here is advice on scoring or interpretation. This is about the software: what it can get wrong with numbers, what we check, and what checking can't do.

Every score in a report came from somewhere in a stack like this. The question is whether the draft can prove it.
Why is AI unreliable with numbers?
Because a language model doesn't calculate or look things up when it writes. It produces the next most likely piece of text, and to the model a number is just more text.
That's fine for prose. It's a problem for scores. In a sentence about a child's reasoning skills, "a standard score of 104" and "a standard score of 114" are equally fluent. Nothing in the way the model writes makes the true one more likely than the wrong one, unless something outside the model holds it to the source.
Assessment files make this harder. A score report puts several numbers side by side for every subtest: a raw score, a scaled or standard score, a percentile, a confidence interval, sometimes an age equivalent. A file may also hold a previous assessment with the same tests and different numbers, plus questionnaires from three informants laid out in near-identical tables. Different publishers format their tables in ways that look alike at a glance. The ways to pick the wrong number multiply fast.
And the error doesn't look like an error. A transposed score reads exactly like a correct one. You'd catch a garbled sentence in a second. You might not catch a 9 that should have been an 8.
What does "checked against the source" actually cover?
Every score written into the draft is cross-checked against the source documents you uploaded. If the draft says a score is 92, that 92 has to be found in the file, attached to the right test and the right subtest.
What it is designed to prevent:
- No numbers from nowhere. A score in the draft should be traceable to the file, not to the model's sense of what a typical score looks like.
- The right column. A percentile shouldn't turn up where a standard score belongs.
- The right assessment. When the file contains a previous report, the scores in the current results should come from the current testing.
- Consistency across the report. The same score shouldn't appear as one number in the results and another in the summary twenty pages later.
We recognize around 199 instruments, and a lot of this year's work has gone into telling similar-looking ones apart. That work is unglamorous and never finished; publishers release new editions and lay out their reports differently. But it's the foundation. If the number is wrong, the paragraph built on it is wrong too, however good it sounds.
Why should the label come from a table, not the AI?
Because the descriptive label next to a score is a lookup, not a judgment the model should be improvising.
When a report pairs a score with a classification word, that word comes from a defined band. If you let a language model write it, you're relying on its memory of how bands usually work, and bands differ between tests. So the label in our drafts comes from fixed tables, not from the AI. The model writes the sentence around it. It doesn't choose the word.
That keeps the draft predictable. The same score gets the same label every time, from the same table, and it matches what the score report says. What that label means for this particular person, and whether it's the right frame for the referral question, is still yours. I've written about how psychologists turn numbers into sentences in from raw scores to report narrative, and none of that changes here.

To a language model, a score is just more text. Something outside the model has to hold it to the file.
What doesn't score checking catch?
This is the part I'd want to know if I were you, so here it is plainly.
It can't catch an error that's already in the file. If a protocol was mis-scored, a score report has a typo, or the wrong birthdate meant the wrong norms were used, checking against the source will faithfully confirm the wrong number. The check proves the draft matches the file. It can't prove the file matches reality.
It can't check what isn't there. If a score report is missing from the upload, there's nothing to check against. The right behaviour is for the draft to leave the gap visible rather than fill it, and a gap is something you should look for on your review.
It doesn't check the words. Score checking covers numbers and their labels. Qualitative observations, interview content and the narrative around the scores are drafted from your material and need your read like any other prose.
It doesn't decide which scores matter. Which results lead the section, which ones you set aside and why, how you weigh them against everything else — that's the professional work, and it stays with you.
What should you still check yourself?
The checking we built takes the most mechanical risk off the table. It doesn't take you out of the loop, and it shouldn't. A short routine still makes sense:
- Spot-check a handful of scores against the score report, especially in the summary, where numbers get restated.
- Confirm the identifying details that drive norms: age at testing, grade, test dates.
- Look for anything missing. If you administered it and it isn't in the results, find out why.
- Read the interpretation as if you wrote it from scratch, because your signature says you did.
Ontario's college put this well in its 2024 Standards of Professional Conduct: technology may not be used instead of a registrant's own professional knowledge. Checking scores is a safeguard on the draft. It isn't a substitute for the person signing it.
Why did we build it this way?
Because of what one wrong number does to a review.
If a psychologist finds a single transposed score, the natural response is to stop trusting every number around it, and the review turns into re-checking the entire results section by hand. At that point the draft has cost them time instead of saving it. That's exactly the wrong outcome for someone whose report already takes ten to fifteen hours to write.
So we treated the numbers as the part that has to be boringly, verifiably right, and let the drafting effort go into the prose. A psychologist who uses Soma described the experience of giving it everything:
"I've thrown the whole kitchen sink at it — my interview notes, the collateral documents, the test data — and it pulls the report together."
Pulling it together is the drafting. Making sure the test data in the draft is the test data in the file is the checking. You need both, and you need to know which one you're relying on.
If you'd like to see it on your own material, bring one of your past reports with identifiers removed and we'll show you the draft it produces, scores and all. You can book that here. You review every line and sign it, as always.
Thank you for the care that goes into getting every number right. We're trying to deserve it.
— Ian
