Most Anki rubrics run too late. Move them upstream of emission.

A rubric is a small set of named criteria that decides whether a question is good. The same word covers two very different uses. The first is post-hoc review: a checklist a student runs on already-generated cards (factual correctness, clarity, distractor quality, grammar parallelism). The second is generation-time gating: the same criteria, but applied as constraints inside the generator before each card is emitted. The two share criteria but differ in where the work happens. Post-hoc means the bad cards exist and you delete them; in-flight means the bad cards are never written.

Three reasons. One, deletion is more expensive than non-emission, especially at deck scale (200 cards per lecture). Two, a post-hoc rubric assumes the reviewer is fluent in the source material; a tired med student grading their own cards at midnight misses pretrained-drift errors that an in-flight check would have caught against the actual upload. Three, post-hoc rubrics implicitly accept the field-average hit rate (roughly 67.9 on the four-criterion held-out eval) because they cannot create cards that did not get generated; in-flight gates create the option of a card that simply does not get written until the constraint is satisfied.

Source-anchoring (the correct answer must trace to a span in the uploaded slide deck, PDF, or transcript). Length-matching (distractor option lengths must be within roughly 25 percent of the correct answer's length). Filler-template ban (no card can ship with 'all of the above', 'none of the above', 'both A and B', 'it depends' as a candidate option). Grammar parallelism (the stem and each option must parse together: singular-singular, plural-plural, 'a' before consonants and 'an' before vowels). The four map onto the four-criterion held-out eval at studyly.io/quality.

Mostly, with caveats. The grammar and length checks transfer cleanly: tell the model to count characters and abort if any option is more than 25 percent longer than the correct answer, and to verify singular/plural and article agreement before emitting. The filler ban transfers cleanly: instruct the model to never emit 'all of the above', 'none of the above', etc. Source-anchoring is the one that does not transfer cleanly: ChatGPT cannot easily verify that the cited span exists in your upload, so the gate degrades to a self-report and the model will sometimes lie about which slide a fact came from. A prompt template that gets you ~70 percent of the way there is in the section below.

The distractor rubric at studyly.io/t/anki-card-distractor-quality is a post-hoc grading instrument: five checks you run on twenty existing cards, with the source slide open in another tab, taking 90 seconds per card. This page is the upstream version: the same dimensions reorganized as gating constraints applied DURING generation, and a discussion of why moving the rubric upstream changes the field-average outcome. Distractor quality is the single criterion with the widest tool-to-tool spread; generation-time gating is the structural reason the spread exists.

On a typical 90-slide cardiology deck, roughly 15 to 20 percent of the model's first-draft cards do not pass all four in-flight checks. The retry budget regenerates the failing cards (different distractor pool, different stem phrasing) until the gates pass or the budget exhausts. The cards a user actually sees are the ones that cleared the gates. The deck arrives with about 200 surviving cards, not 240 cards 40 of which are quietly bad.

The NBME Item-Writing Guide is the authoritative source for the underlying principles: distractors should be plausible and parallel, filler shapes are item-writing flaws, grammar mismatches give away the answer. The guide is written for human item-writers preparing licensure exams. It does not specify ±25 percent length tolerances or 'gate per card before emission' wording because those are implementation choices for an automated generator. The page below ports the NBME principles into operational gates that a generator can actually run.

Yes, but at the deck level not the card level. A type-coverage gate looks at the running tally of MCQ vs free-response vs case-style vs image-occlusion across the deck and biases the next emission toward whichever bucket is underrepresented relative to the target mix. This is one level above the per-card rubric, which is why it is not in the four-checks list. It is the reason a Studyly export of a single lecture has a balanced question-type mix instead of 200 single-best-answer MCQs.

Marginally. Each gate adds a small amount of work per card: a substring search for source-anchoring, a length comparison for length-matching, a regex for filler templates, and a quick grammar parse for parallelism. The 60-second-per-90-slide-deck number on Studyly is measured with all four gates active. The throughput cost is real but small enough that it is dominated by the generation step itself.

Yes, at studyly.io/quality. The leaderboard breaks the held-out three-document eval down by criterion: factual correctness, clarity, distractor quality, question-type coverage. Studyly 81.3, Unattle 78.0, Gauntlet 68.0, Turbolearn 57.8. Distractor quality is the criterion with the largest tool-to-tool spread, which is consistent with the in-flight-vs-post-hoc framing on this page: tools that gate on distractor quality at emission time pull ahead of tools that don't.