Work · IDEAcademy · Perth · February 2026 onward

A school asked AI to help write three years of curriculum for teenagers. Here is how we made sure it could not make things up without someone noticing.

IDEAcademy is a three-year alternative to Years 10 to 12 for learners aged 14 to 18, designed for a substantial neurodivergent cohort, working toward nationally recognised vocational outcomes. A wrong sentence in a masterclass reaches a young person.

A ski resort on Mars · Soren's illustration

Rebecca designed the educational authority, learner model and acceptance rules; Soren designed the LLM Contract and engineering system that operationalised them; the deployed LMS Compiler is the result of their collaboration.

The method

Sources are ranked before the model runs. Code builds the authoritative shell. The model may fill only declared fields. Protected fields are fingerprinted. A forged title is caught as a hard failure. A person promotes to production.

The demonstration film · 0:25 to 2:30 · sound on
  • 116 masterclass bundles | 3 production courses | April 2026 scoped run
  • 494 companion files | performance tasks held behind gates
  • Zero reported deployment failures in this run
116 masterclass bundles
3 production courses · April 2026 scoped run
494 companion files
performance tasks held behind gates · April 2026
11,380 image references rewritten
1,121 pages and assignments · May 2026 · zero reported failures in that run, four sources unresolved
363 verified session-plan specifications
at 30 June 2026 · 'verified' is a pipeline state, not universal human approval
6,190 student-quiz submission instances
June 2026 read-only scan · activity, not unique learners

Where it went wrong

A human caught the model.

In April 2026 review found an AI session had fabricated or paraphrased part of a BSBTEC301 performance-evidence requirement. The error was checked against the external authority and the source-authority rule was strengthened. Oversight working, not system perfection.

A fix that did not hold.

In May 2026 answer-length bias was found in 11 masterclasses. A shared helper, a compile-time gate and 14 tests followed, and 28 masterclasses were rebalanced. Later regeneration exposed recurrence. We do not claim permanent resolution.

What it became

What the engagement produced became the products. The contract that bounds what a model may touch became the rules in Forge. The weekly digest and findings tracker a second operator works in every Monday became the thread the Bridge holds. The rule that a different model family must sign a review became Quorum.

The LLM Contract was developed by Soren Tellegen through Digital Pulse while consulting to IDEAcademy.

Building something where a wrong answer costs something?