Why I built REHEARSE
REHEARSE began with a frustration I know from workshop training: reading the correct procedure is not the same as making a good decision when the evidence is incomplete. A learner can remember an answer and still replace the wrong part with complete confidence.
That is the moment I wanted to make practiceable.
Textbooks tell. Videos show. REHEARSE lets you practice.
What it is
REHEARSE turns expert source material into an interactive, evidence-grounded simulation. It is not a quiz generator and it is not a generic chatbot. The engine compiles knowledge into a causal scenario graph, lets a learner make decisions with consequences, diagnoses the misconception behind the reasoning, and gives an instructor evidence of how the learner thought.
I chose one deep automotive case because it makes the idea easy to test. The product itself is a reusable education engine, not a car enthusiast app.
The case a judge can play
The learner is the diagnostic technician for a fully synthetic vehicle called the Aster Atlas. The customer reports hesitation and a warning that return in wet conditions. A generic range/performance fault appears to implicate a three-wire, 5 V analogue airflow sensor.
Replacing that sensor is plausible, but premature.
If the learner anchors on the fault code and replaces it, the simulation adds 45 minutes and $280 in illustrative cost. The original fault remains. REHEARSE does not simply mark the choice wrong: it shows what changed, what did not change, and which evidence the hypothesis failed to explain.
The tutor identifies component anchoring and asks a targeted Socratic question instead of revealing the answer. The learner can recover by testing the circuit under the reproduced condition. The system map and measurement cards make the causal relationship visible:
0.93 V local sensor output + 0.35 V ground offset = 1.28 V observed by the control unit.
The signal is electrically possible, but outside the expected synthetic operating envelope. A loaded voltage-drop test and controlled connector test localize a high-resistance ground connection. After the connection is repaired, the learner remeasures under the same condition, verifies the signal at 0.96 V, and confirms that the complaint does not return.
Every value, vehicle, price and time estimate in this case is synthetic. The case uses no BMW specifications or proprietary manufacturer data. It is an educational simulation, not authorization or instruction for a real repair; real work must follow applicable service information and qualified professional judgment.
One scenario, three useful views
Expert / author
An expert can paste structured source material, have GPT-5.6 compile a scenario candidate, inspect source provenance, and see sourced facts separated from inferences and uncertainties. A deterministic validator checks schema, reachability, missing evidence, contradictory states and unsafe or unsupported claims. Nothing is publishable until an expert reviews and approves it.
Learner
The learner receives a work order, confirms the symptom, inspects, measures, forms a hypothesis and verifies the repair. Evidence appears only when it has been discovered. Wrong decisions create recoverable consequences in time, cost and unresolved symptoms. Completion produces a competency result, not just a quiz score.
Instructor
The instructor sees the decision path, observations requested, evidence used or ignored, hypotheses, unnecessary actions, simulated cost, diagnosed misconception, confidence, uncertainty and recommended next practice. That makes reasoning visible, including where a learner recovered.
Where GPT-5.6 is genuinely used
GPT-5.6 has two focused jobs:
- Compile untrusted expert prose into a provenance-bearing scenario candidate using Structured Outputs.
- Phrase a misconception-specific Socratic question from tightly constrained synthetic context.
It does not control scenario state, costs, scoring, publication or repair authorization. Those boundaries stay deterministic and testable.
The public judge experience is intentionally keyless and deterministic so it remains fast and reliable. Separately, I ran a sanitized owner proof against the exact gpt-5.6-sol model. The tutor call succeeded. The compiler response crossed the model boundary but was rejected by the product's deterministic validation because it did not satisfy every publication rule. I kept that rejection in the evidence instead of presenting it as a success. For me, that is an important part of the product: model output is a candidate, never authority.
Approved scenarios are cached, and model calls are reserved for generation and adaptive wording rather than ordinary learner transitions.
How Codex accelerated the build
The majority of the project was created in one primary Codex build thread, starting from an empty, isolated repository. Codex helped me move from concept to a working vertical slice: TypeScript scenario contracts, validator, deterministic state engine, author workflow, learner simulation, instructor analytics, API boundaries, security controls, tests, deployment records and the automated demo-video pipeline.
Codex was especially valuable for keeping implementation, tests and release evidence moving together. It also ran bounded review passes against the published judging criteria: technological implementation, design, potential impact and quality of the idea.
The important human decisions remained mine: the problem to solve, the choice to build one complete scenario rather than several shallow demos, the workshop failure mode worth teaching, the evidence sequence, the safety boundary, and what the AI is and is not allowed to decide. The built-in synthetic case was then reviewed and corrected from real human automotive workshop experience.
Primary Codex /feedback Session ID: 019f7a57-0f25-7783-8089-1ffeb8d7da46
Technical implementation and verification
REHEARSE is built with Next.js, React, TypeScript, Zod, the OpenAI Responses API with Structured Outputs, OpenNext and Cloudflare Workers.
The release includes:
- 45 deterministic unit and integration tests
- 12 desktop/mobile development browser checks
- 6 optimized-production browser checks
- 12 public-host checks
- keyboard and responsive interaction checks
- input and output validation, origin protection, rate limits and a global kill switch for live mode
- prompt-injection defenses that treat uploaded material only as data
- dependency, license, accessibility and secret-leak audits
- a private repository with an immutable submission tag and integrity manifest
A second generic laboratory fixture in the test suite verifies that the engine is not hard-coded to automotive vocabulary. Automotive diagnostics is the demonstrated vertical; the underlying pattern applies wherever people must interpret evidence, make consequential decisions and learn from how they reason.
Potential impact
The first audience is vocational instructors and workshop onboarding teams. REHEARSE could help them practice rare, costly or unsafe-to-recreate decisions before those decisions happen on a real asset.
A pilot should measure whether learners choose more discriminating tests, reduce unnecessary simulated cost and time, recover from misconceptions, and need less instructor intervention on repeated cases. Those are prospective evaluation goals, not claims of adoption or proven learning gains.
Don't just study the decision. Rehearse it.
Built With
- axe
- codex
- gpt-5.6-sol
- next.js
- openai-responses-api
- opennext
- playwright
- react
- structured-outputs
- typescript
- vitest
- zod