EduOS

Inspiration

12 sources. 2 bounded options. 1 accountable teacher decision.

AI can generate a lesson in seconds. The harder professional question comes first: which strategy is best supported for this class, this learning objective, and this moment?

The information needed to answer that question is often distributed across records, observations, learning plans, prior activities, and pedagogical guidance. Reconstructing that changing context consumes attention before the teacher can even begin the creative work of teaching. EduOS explores a different role for AI: not another content generator, but a governed decision-support layer that makes evidence, uncertainty, and professional responsibility visible.

Its intended human value extends beyond efficiency. EduOS is designed to reduce avoidable cognitive friction in preparation and preserve more teacher attention for teaching, relationships, reflection, and life beyond administrative work. That is a product hypothesis to validate, not an outcome claimed by this prototype.

What it does

The demonstrator uses one fixed synthetic secondary-school class dossier, 12 sources, two bounded options, and one decision question. A teacher invokes the EduOS skill in Codex with GPT-5.6 Terra. The skill reads the complete dossier, distinguishes recorded facts, professional observations, external guidance, and system inferences, and prepares a cited strategy brief.

The brief shows the option currently best supported by the available evidence, the reasons for that recommendation, its limiting evidence, information gaps, uncertainty, and the alternative of deferring until missing constraints are checked. The teacher then remains free to adopt, adapt, reject, or defer. EduOS does not turn its recommendation into the teacher's decision.

Why EduOS is innovative

The innovation demonstrated here is the sequence and control structure around the model:

• Decision before generation: EduOS supports the professional choice that should precede lesson creation.
• Evidence before fluency: material claims must resolve to known sources, not merely sound plausible.
• Validation before presentation: a deterministic gate rejects non-compliant output before the interface can present it as complete.
• Human authority after recommendation: the system's recommendation and the teacher's final choice remain visibly separate.

A conventional prompt-and-response workflow can end with a fluent answer. EduOS adds a bounded source set, explicit epistemic types, closed options, visible uncertainty, machine-checkable constraints, and a human decision boundary. The result is an inspectable decision artifact rather than an opaque answer or automatically generated lesson.

Product experience

1. The teacher sees the decision, the synthetic context, and the source inventory.
2. GPT-5.6 Terra synthesises only the bundled evidence through the repository skill.
3. The validator checks the brief and binds it to the case with SHA-256 hashes.
4. The browser companion opens that exact validated artifact and makes every citation inspectable.
5. The teacher considers the recommendation and records no choice unless they explicitly decide to do so.

The experience is deliberately complete at one narrow vertical: understand the context, support a decision, explain the recommendation, and preserve human control.

Technical implementation

Synthetic dossier -> Codex skill + GPT-5.6 Terra -> strict JSON brief -> deterministic validator -> hash-bound report -> browser companion -> teacher choice

The repository-local skill contains the instructions, fixed case, strict JSON Schema, and dependency-free Python validator. The validator rejects unknown citations or options, changed option text, invalid epistemic classifications, inconsistent output states, learner scoring or ranking fields, diagnostic inference, guaranteed-outcome language, and artifacts changed after validation.

The browser companion uses dependency-free Python, HTML, CSS, and JavaScript. In artifact mode it makes no model call: it displays the generated brief only when the report and the current case and brief hashes match. Twenty-two deterministic tests exercise the contract, failure behaviour, artifact integrity, and human-control boundary. Model identity is demonstrated by the visible Codex host state; the validator truthfully proves only contract and grounding checks.

How I built it with Codex

I defined the EduOS product vision, selected the professional decision, supplied the scenario evidence, and approved the boundaries. I used Codex as my implementation collaborator to turn those decisions into a governed repository, trace requirements across the Canon, Product Bible, functional specification, architecture, data model, UX, prompt contract, and tests, and implement the working vertical.

Codex accelerated the translation from product intent to executable constraints. Key decisions included narrowing the submission to one complete decision, using a repository-local Codex skill as the accepted GPT-5.6 runtime, separating model evidence from what deterministic validation can prove, and introducing a hash-verified handoff between the generated brief and the interface.

Potential impact on teachers' work and quality of life

For teachers, quality of work and quality of life are connected. EduOS is designed to reduce the avoidable effort of repeatedly reconstructing context, keep the basis of a decision reviewable, and protect professional attention for work that requires human presence and judgment.

If validated in practice, this approach could reduce time spent searching and reconciling fragmented information, make preparation less cognitively demanding, and leave more attention for learners and life outside administrative tasks. This demonstrator establishes a working, testable approach; it does not claim measured time savings, reduced workload, improved wellbeing, or improved learning outcomes.

Trust boundary

The project uses no real learner data and does not diagnose, rank, grade, profile, or predict learners. It does not generate a lesson, execute a strategy, or claim that the recommended option will be the most effective. External guidance informs the brief but never overrides local evidence or teacher judgment.

Accomplishments

• One decision is traced from product doctrine to executable acceptance tests.
• Twelve heterogeneous sources are classified and made inspectable.
• Two bounded options are compared without allowing invented alternatives.
• Every material recommendation claim is linked to known source identifiers.
• The generated brief is validated and cryptographically bound to its case before presentation.
• Twenty-two deterministic tests cover both the successful path and unsafe or unsupported outputs.
• The primary judging path requires no package build, separately funded API request, or real learner data.

What I learned

A trustworthy decision-support experience needs explicit boundaries as much as model capability. The strongest signal is not fluency; it is whether the system can show what it knows, where it came from, what remains uncertain, what the validation actually proves, and who retains authority.

What is next

The next step is validation, not feature expansion: co-design with teachers, evaluate decision clarity and time-to-decision, examine perceived cognitive burden, and test whether provenance and uncertainty are genuinely useful in practice. Any move toward real educational use would also require institutional and legal review, privacy and security architecture, accessibility evaluation, and explicit approval before processing real data.

Built With

• codex
• css
• gpt-5.6-terra
• html
• javascript
• json-schema
• python