Draft questions come in from the generator. A council of frontier models judges every one for correctness, ambiguity and source support. Only unanimous passes ship to vetready.ca — the rest land in a reviewer's queue.
The back of house for vetready.ca ↗ — not a public product.
A 12-year-old Cavalier King Charles Spaniel presents with a grade IV/VI systolic murmur loudest at the left apex. Radiographs show left atrial enlargement. Most likely diagnosis?
Every draft is judged on the dimensions that actually break exam questions — then routed by agreement, not by vibes. A full council on the base pass; a frontier adjudicator to settle any split.
Each reviewer decides whether the keyed answer is clinically correct and whether a second option is defensibly right. Split verdicts route straight to a human.
An explanation can't attribute claims the cited material doesn't contain. Questions that fail source-support are flagged, not shipped.
If an image is AI-generated on a diagnostic item, a deterministic rule fails it outright — the answer can't rest on a picture that was never verified against ground truth.
Auto-approve fires only when every reviewer passes with no flags. Any disagreement becomes a pending item with the pass/fail split attached.
Reviewers refine the stem, options or explanation and send it back through the council. Rule-based image fails resolve instantly, so the loop never spins on a question that can't pass.
One verdict row per model per question, plus the deterministic gate kept for the trail. You can see who passed, who failed, and why.
Every draft moves through the same disciplined path, from raw source material to a published, board-ready question. Nothing skips a step, and nothing ships until it's earned approval.
Every reviewer on the council judges the base pass independently — four verdicts per question, one row each, no model seeing another's answer.
Unanimous pass, no flags → auto-approved. Any pass/fail split promotes the question into the review queue with the count attached.
A flagged text dimension escalates to a mid or frontier adjudicator — web search on only for source-support and factual claims.
Approved questions publish to vetready.ca ↗, the consumer product students actually study on.
If it comes from the machine, it's in mono. Reviewers never read model names in Simple mode — just the plain-English suggestion and three buttons.
Access is redeemed with a single-use code, not a public form. Reviewers see the pending queue; admins see the pipeline. Every write goes through a server function — a direct API insert from a normal account fails by design.
No signup path exists. A tiered invite code attaches your membership; used, expired or email-locked codes are rejected.
Reviewer, senior and admin roles are checked in RLS, not just the UI. Lower tiers get zero rows on restricted tables even via direct calls.
Doctors get a plain-English bench and big Approve / Reject / Needs-edits. Admins get the full technical panel — per-model verdicts, cost, tokens.
vetready.click is invite-only. Redeem a code to review the questions that become vetready.ca — or ask for one if you're a veterinarian who wants to help.
Source of VetReady.ca ↗