Technical page

How NexGen Mastery scores, selects and calibrates items. This is the page a faculty member or a careful student should be able to find before trusting a readiness band. Every constant below is read from the engine code and every count from the attempts table; the live copy is at nexgen.myomedprep.org/#/technical (GET /api/technical).

Last updated: October 2026. Read alongside the Trust page, the coverage map and the evidence standard.

No psychometrician has signed off on the model, the thresholds, or this page. Item difficulties are authored ratings placed on a logit scale, not parameters estimated from response data; the passing standard is the bank's own centre and is not linked to any NCSBN, ANCC, AANP or ARRT standard. A readiness band is a statement about performance on our bank — not a prediction of a licensure or certification result.

Item statistics from real attempts

For every scored item we compute, from learners' first attempts only:

Exposure (n) — how many distinct learners have met the item.
Difficulty (p) — the mean fraction of the item's marks earned, so partial-credit NGN items are handled the same way as single-answer items.
Discrimination (r_pb) — the corrected item-rest point-biserial: the correlation between a learner's score on the item and their mean score on every other item they have answered.
Median time — the median of client-reported seconds from the item being shown to the answer being submitted.
Distractor analysis — for multiple-choice and select-all items, how often each option is chosen.

An item is calibrated once 30 learners have met it; before that it is provisional and is not judged. Calibrated items with p above 0.95, p below 0.15, r_pb below 0.10, or negative r_pb are flagged for human review. Nothing is retired by the software. Statistics are recomputed nightly and the staff item table and CSV export are limited to the operator; the public sees bank-level aggregates.

Snapshot at publication (October 2026): 9,982 items in the bank (including 833 MyoMed Rescue EMS items in draft pending medical-director review), 0 calibrated. Attempt timing began being recorded with this release, so the first calibrated items and the first median times will appear as learners work through the bank — the live figures are on the NexGen page linked above.

Adaptive practice

Measurement model	One-parameter logistic (Rasch) with an Elo-style running ability update, K = 0.40.
Item difficulty	Authored rating 1–5 placed at −1.6, −0.8, 0.0, +0.8, +1.6 logits.
Ability range	−3.0 to +3.0 logits.
Item selection	The unseen item whose difficulty is closest to the learner's current ability.
Exposure control	One of the three closest unseen items is chosen at random, so two learners at the same ability do not see the same sequence.
Recycling	Once every item in scope has been seen, items are re-served and marked as repeats; repeat attempts are excluded from item statistics.

Readiness examination

Length	85–150 items, of which 15 are unscored pretest items.
Time limit	300 minutes.
Ability estimate	Newton–Raphson maximum likelihood, refit over the whole response history after each item; starts at 0.0, bounded −4.0 to +4.0.
Confidence interval	95% (z = 1.96). The standard error is never reported below 0.18, so the interval is never narrower than ±0.35 logits.
Passing standard	0.0 logits — the bank's own centre.
Item selection	The Client Need category furthest behind its published share is served first, then the most informative unseen item in it (random among the three closest).
Clinical judgment floor	At least 10% of items are clinical-judgment formats (bow-tie, matrix, cloze, highlight, ordered response); the result reports the mix that actually occurred.

Stopping rules

300 minutes elapsed.
150 items delivered.
After 85 items, the lower bound of the 95% confidence interval is above the passing standard.
After 85 items, the upper bound of the 95% confidence interval is below the passing standard.

Blueprint weights

The midpoint of each NCSBN 2026 Client Needs band: Management of Care 18%, Safety and Infection Control 13%, Health Promotion and Maintenance 9%, Psychosocial Integrity 9%, Basic Care and Comfort 9%, Pharmacological and Parenteral Therapies 16%, Reduction of Risk Potential 12%, Physiological Adaptation 14%. How the bank itself sits against those bands is on the coverage map.

Readiness bands

Very high	Lower bound of the confidence interval more than 0.75 logits above the standard.
High	Lower bound above the standard.
Borderline	Interval straddles the standard.
Low	Upper bound below the standard.

Synapsis accuracy benchmark

Synapsis, the on-site tutor, is run against a fixed prompt set every time its model, prompt or retrieval changes: medication safety, contraindications, doses, paediatric and pregnancy cases, misleading premises, hypothetical clinical vignettes, out-of-corpus questions, and messages it must decline (real-time bedside care, personal medical advice, crisis, graded work). Each answer path — the lesson-text extractive path, the on-site model and the on-device WebGPU model — is graded separately, after the same citation and evidence gates a learner's turn passes through, on citation accuracy, unsupported-claim rate, dose-hallucination rate, refusal accuracy and misleading-prompt resistance. Every published number carries the model name, the date and the commit it was run at.

The dated results table is on the live technical page at nexgen.myomedprep.org/#/technical (GET /api/technical/synapsis). A path that has not been run for the current model is shown there as “not yet run”, never as a number.

This is a regression gate the team runs, not a clinical-safety certification and not a promise about any individual answer. The prompt set is small and written by us; its lexical checks can pass a wrong answer and fail a right one worded differently. Every Synapsis turn still shows its own evidence state to the learner.

Limitations

Item difficulties are authored ratings, not estimated parameters. Calibration from real attempts is what the item statistics above are collecting towards; until an item is calibrated its difficulty is an author's opinion.
The passing standard is not equated to any licensing or certification body's standard, and we publish no pass-rate prediction.
Learners who have answered only one item contribute to p but not to discrimination.
Reported time is what the learner's browser measured; a tab left open counts up to an hour. The median is used because it tolerates that.
Calibration thresholds (n = 30; p 0.15–0.95; r_pb 0.10) are our working defaults pending psychometric review.

Live technical page Outcomes Coverage map Trust & privacy