How CorrectMED AI is used

Who does what

• CorrectMED AI reads. It copies what’s printed on a delivery note, on a pack and in a NAFDAC notice, marks every value as clear, unclear or absent, and boxes the part of the photo each value came from. It never corrects, completes or guesses, and its answers have no field for a judgement about a medicine.
• Fixed rules compare. Code, not AI, compares what was read with the delivery note, the counts your team enters and the published NAFDAC notices, and picks one result per line.
• People decide. A receiver counts. An admin or a pharmacist accepts, part accepts, quarantines or rejects each line and gives a reason. A viewer can only look.

Text in a photo is treated as data, never as instructions, and the reader is never told what a value is expected to be.

When it can't read something

• A value it can’t read says Could not read, and never reaches the rules.
• A value with no box says “source region unavailable”.
• A photo that isn’t a medicine pack says “Not a medicine pack, nothing read”.
• If reading fails, the page tries again, then shows Try again. A failed read never gives a result and costs nothing.
• The same photo uploaded twice in a workspace is read once and labelled as read earlier.

The fixed rules

• Only clear values are compared. Unclear and absent are kept apart.
• Quantities are compared in packs, through packs per carton and units per pack. Quantity is a staff entered count with an audit trail. Per unit scanning and aggregation codes are not supported.
• Strength, form and pack size all count when a pack photo is linked to a line.
• Batch and expiry on the pack are compared with the note.
• Short dated means under 182 days to expiry.
• A date that can be read two ways keeps both readings, and the earlier one drives the expiry check.
• A photo already used on another receipt is flagged.
• NAFDAC notices match as exact (a reviewed notice with the same batch, product and strength) or uncertain (for example a batch one character off, or a notice that hasn’t been reviewed). Uncertain matches show in amber.

Each line gets the first of these results that applies:

1. Published alert match
2. Published alert match (uncertain), review required
3. Unable to assess, with the reason
4. Reference unavailable, when there’s no approved packaging reference to compare with

Differences in quantity, batch or expiry are findings beside the result; they never change it. A visual match, or finding no alert, is not proof that a medicine is genuine or of good quality.

The report

Every receipt prints the same fixed template, never free text written by the AI:

• the audit ID, and the Sandbox mark in a sandbox workspace;
• the photos with their source boxes, or “Photo deleted under the retention policy”;
• each value with the photo it came from, and any correction with who made it and when;
• the quantity on the note and the quantity your team counted, kept apart;
• the findings, each result with the limits line, and the state of the NAFDAC copy at the time;
• each decision with the name, role, reason and time.

You can print it from the browser or export it as a CSV.

How long it's kept

• Sandbox photos: 30 days. Sandbox receipts: 90 days after their last activity.
• Other photos: 12 months. Receiving records: 6 years, or until you ask us to delete them.
• Your photos never become packaging references, and they aren’t shared with other workspaces.

Where it runs

Records and photos

Google Cloud, in its EU multi region.

Reading

Google Cloud, through Google's European endpoint.

Sign in

Google’s global service.

Hosting

Vercel, in Frankfurt, Germany, with request logs kept globally.

The reader

The reader is called cm-reader-1. Reviewers can ask us for the exact model version behind it, and the evaluation shows how well it reads. You can try it yourself on the sample notes.