Your LMS tracks completion. Attest tracks competence.

SCORM reports "Completed." Auditors want mastery evidence. A senior BSA analyst who already knows the material and a six-month branch manager with real gaps sit through the identical 4-hour AML course and earn the identical checkmark. The record cannot tell them apart, because it was never measuring what either one knows.

Incumbents that market themselves as "adaptive" mostly run collaborative filtering ("people like you took Course X next"), not concept-level knowledge tracing ("you have not mastered concept Y"). And employees increasingly use AI assistants to breeze through modules, so the completion record says "trained" when the learning never stuck.

In regulated work that gap is a liability an auditor can price. U.S. corporate training spends about $102.8B a year at roughly $874 per learner (Training Magazine, 2025), and much of it buys a completion record rather than proof of competence. Where a training gap carries regulatory consequence, in AML and BSA recertification or under EU AI Act Article 4 role-based AI-literacy rules that begin August 2, 2026, a green checkmark is exactly the wrong evidence to bring.

The Senior BSA Analyst (8 years in role) reads green almost everywhere, weak only on Trade-Based Money Laundering and one SAR red-flag concept. Attest certifies 16 of 18 concepts, routes 2 to needs-proof, and does not flag the evidence. Seat time drops from 240 minutes flat to 47.3 minutes adaptive, 80.3% saved. This is the "already knows most of it, skip what is proven" case.

The New Branch Manager (6 months post-transfer) has real gaps across SAR narrative and TBML. The model infers low mastery on several concepts and routes them to flow or scaffold rather than skip. Attest certifies 10 of 18 concepts, routes 8 to needs-proof, and cuts seat time from 240 to 146.7 minutes, 38.9% saved. Fewer concepts certified than the analyst, because this learner genuinely needs the training. The savings track the mastery, not a flat discount.

The Operations Associate is the hard case. This learner answered fast and perfect across 18 unrelated concepts, a response-time distribution inconsistent with genuine recall. Apparent mastery looks high, but the competence gate flags the evidence and certifies 0 of 18. All 18 route to needs-proof, evidence is marked flagged, and seat time is withheld pending verification. We frame this as value, not a "gotcha": we only certify what we can prove.

The cohort result, computed live

Over a genuinely-simulated 500-person annual-recertification cohort at a $75 per hour fully-loaded wage, Attest computes about 51.9% cohort seat-time reduction (2,000 flat hours down to roughly 962 adaptive hours, about 1,038 hours recovered), about $77,872 recovered on this one module, and about $389,362 annualized when scaled to a typical 20-hour compliance program. The dollar figures are computed from the model's outputs, and the annualized number is labeled a projection on screen.

The model benchmark, the independent anchor

Held-out AUC is 0.8315 on a seeded, reproducible synthetic PFA-generated KT corpus (2,000 train and 500 test students, 18 concepts, 12 epochs), with next-step accuracy 0.771. Because the data-generating process (Performance Factors Analysis, a classical logistic learner model) is a different model family from the transformer we train, the held-out AUC shows SAKT recovered real latent learning structure rather than grading its own assumptions. For external context, published SAKT AUC on the public ASSISTments dataset is about 0.80 (Pandey & Karypis, 2019). The durable headline stays the percentage of seat time saved and the percentage auto-certified, which hold at any model quality. The AUC proves the model is real knowledge tracing, not the product's headline number.