Pinnacle runs on a knowledge graph of 386 skill nodes across seven certifications — each one verified against the awarding body's own published outline — with Bayesian mastery tracking on top and an AI voice tutor whose every session traces back to named learning outcomes. This page is the engine, opened up for the people who steward those syllabi.
No account needed, no model spend per visitor — the same 25-question diagnostic your members would take.
Most prep providers write a table of contents and call it coverage. We model each certification as a graph of skill nodes with prerequisite edges — and in August 2026 we verified every node's name and scope against the awarding body's own published curriculum documents, reading the primary sources in full rather than summaries of them. Nodes teaching retired material were retired. Modules with no node were added. Descriptions that had drifted from their outcomes were rewritten from the outline's own words.
The result is a curriculum spine that says, for every concept, exactly which published learning outcome it serves — and refuses to teach anything that serves none.
On top of the graph sits a knowledge-state engine: every answer a candidate gives updates a Bayesian estimate of mastery on the nodes that question actually tests — and only those nodes. Question-to-node mappings are curated and certification-scoped, so practice on one exam never contaminates the estimates of another.
Item response theory and Bayesian knowledge tracing give each skill node a posterior — not a percentage of questions right, but an estimate of what the candidate actually knows, with the item's measured difficulty and discrimination in the model.
The engine tracks how much evidence sits behind every estimate, and the interface is required to say
so. Two answers on a topic produce low credibility — shown as exactly that, never dressed up
as a confident mastery bar. Candidates fail exams on what they were sure about; we refuse to
manufacture that certainty.
The next question is chosen to learn the most about the candidate — not to pad a streak. Combined with FSRS spaced repetition for retention scheduling, study time goes where the posterior is weakest and the exam weight is highest.
"Three models agreeing is not the same as being correct — and a confident bar built on one answer is a lie."
Amber — our voice AI tutor with a live mathematical whiteboard — is bound to the same graph. She opens each session from the candidate's actual mastery state, anchors every thread she teaches to a named outcome, is forbidden from teaching material that serves none, and closes with a concrete practice commitment she will follow up on next session. The session summary opens with the learning outcomes covered — which means a member's study history is an auditable record against the syllabus, not a chat log.
Last session's material, retrieval-checked by name. The practice commitment they made — met, partial, or missed — raised once, with the real numbers.
Each thread names the outcome it serves. Depth follows the outcome's own verb — "describe" is taught differently from "calculate", in both directions.
Candidates commit to a confidence level before hearing whether they're right. Confident-and-wrong is treated as the most important event in the session.
A countable practice target on a named topic by a named day — recorded, and reopened at the start of the next session.
Pinnacle is an independent platform — we are not affiliated with or endorsed by any awarding body. That independence is exactly what makes the verification honest. Here is what we bring to an institutional conversation:
Adaptive preparation as a society or association member benefit — priced for cohorts, with the free diagnostic as a zero-friction entry point your members can take today.
The same engine that tracks mastery for candidates is built for maintaining it across a career — spaced retention scheduling, annually refreshed current-issues content, and outcome-traceable session records that fit CPD evidence requirements.
Exam-track cohorts with real visibility: which outcomes a cohort is weak on, who is on track for their sitting, and where teaching time actually moves the needle — from the engine's posteriors, not self-reports.
The fastest way to evaluate Pinnacle is the way a candidate would meet it: 25 questions, a diagnosis of the root cause rather than a score, and the graph doing its work in front of you.
Institutional enquiries — support@pinnacle-prep.org