A radiology report is four documents at once. That is why this is hard.

1. Picking the right template, from a dictation with no code in it

Report templates are organised by procedure code — hundreds of them, many with near-identical names. A radiologist dictating at speed often supplies no code at all, just an exam header and findings. Match that header by keyword overlap against a hundred similar filenames and you will eventually put a carotid ultrasound worksheet onto a breast template, or report a three-region spine MRI as thoracic only because the first matching file won. Getting this right means resolving codes deterministically, recognising when one study is genuinely several exams, and refusing a match that is anatomically out of field rather than taking the first one that scores.

2. Keeping the radiologist's findings the radiologist's

The machine may organise, format and place. It may not invent, and it may not quietly drop. A confirmatory aside ("right kidney normal") must not shrink the template's full normal line; a new finding on a broad normal statement is added, not substituted; a stated conclusion is adopted, not re-derived. Systems fail here in both directions — hallucinating a paragraph about an organ that was never imaged, or deleting a clause the radiologist dictated — and both failures look like fluent, professional prose.

3. Comparison

"Compared with the prior study" is only useful if the system has the right prior, knows its date, and knows which finding on it is the same finding. Stable baseline attributes established on a prior — a density category, a chronic post-surgical change, a known device — should carry forward without the radiologist re-dictating them, while anything the current study contradicts must not. Priors the radiologist supplied are the source of truth; priors the system found on its own are a fallback, never a substitute.

4. The follow-up recommendation — where the value and the harm both live

An incidental finding with no recommendation is a missed opportunity; one with the wrong interval is a real harm, repeated at scale. The interval is not a matter of opinion: it comes from published criteria that have a version, an applicability boundary, and exclusions that matter (a screening chest CT and a diagnostic one are not governed by the same document). A recommendation that cannot name the criteria and version it followed is not auditable, and in a deposition it is not defensible either. This is exactly what RadVector API returns.

5. Quality assurance that runs on every report, in milliseconds

Laterality that flips between the findings and the impression. A measurement in the wrong unit. An anatomy word that belongs to a different study. A date that cannot be right. These are cheap to catch deterministically and expensive to catch in review, and the check has to be fast enough to run on every report rather than a sample.

6. Documentation integrity for coding

Whether the report documents contrast, laterality, completeness and combined-region coverage decides whether the study is coded and reimbursed correctly. That is a documentation question, answerable from the report against the code's own descriptor — and it belongs next to the clinical check, not in a separate downstream audit weeks later.

This directory is open to other radiology reporting and decision-support systems, including ones that compete with ours. Two rules make that safe to promise:

• A listing never touches an answer. No verdict, recommendation, citation, ranking or reference returned by the API is influenced by whether anyone paid for anything. The decision path has no commercial input, and that is not a policy we can quietly relax — it is why the answers are worth buying.
• Paid placement is labeled as paid. If a listing is sponsored, it says so on the listing, in the same type size as everything else.

Listings are editorial today and free. To be considered, send the system name, a one-paragraph description, the URL, and who to contact.