AEGIS is the generation pipeline behind the product: it researches a goal, designs the arc, writes the lessons, and then tries to break what it wrote before a learner ever sees it.
Current versionAEGIS 2.0Shipped 26 July 2026
What AEGIS is, and what it is not
AEGIS is an engine, not a model. We do not train, fine-tune, or own a language model. AEGIS chooses among third-party models, prompts them, and verifies what comes back.
The version tracks the pipeline, not the models. We change which model writes or judges as better ones appear. That is a configuration change and does not move the engine version. Adding or removing a generation stage or a gate does.
No model grades its own work. The model that writes a lesson and the model that judges it are deliberately from different families, and the independent-assessor tier is never the writer.
Verification is preferred to opinion wherever it is possible. Code is executed, arithmetic is calculated, and citations are resolved against sources actually retrieved. A second model reading the first one is the fallback, not the default.
When a check cannot run, the content ships marked as unverified. It is never presented as verified.
What a course goes through
Every course runs the same path. Each stage can send work back to the one before it.
1
Research
The goal is researched against current sources and the results are stored as a per-goal corpus the writers can retrieve from.
2
Arc design
One pass authors the shape of the course before any lesson exists: where difficulty steps up, what artifact runs through it, which concepts get revisited.
3
Outline
The curriculum is generated against measurable objectives and a knowledge graph, not a topic list.
4
Grounded writing
Lessons, worked problems and quizzes are written against retrieved excerpts. Tool, API and version claims must come from a source or be left out.
5
Execution
Code examples and coding assignments run in a sandbox at generation time. The output a learner reads is the output that ran.
6
Gates
Deterministic fact-checking, citation resolution, quiz linting and a performance floor run over the finished chapter. Failures escalate to a stronger model with the complaint attached, and a draft that fails its judge does not ship.
7
Adversarial read
A long-context judge from a different model family reads the whole free chapter and its quiz together. A simulated learner attempts it. A skeptical-buyer pass asks whether the chapter is worth paying for.
8
Ledger
Every gate outcome is recorded against the course, including passes. Anything that could not be repaired ships visibly marked rather than quietly.
Version history
Dates are when the last mechanism in the release landed in production code.
AEGIS 2.0
The learner does the skill, and the engine can show its work.
The performance spine
The free chapter has to make the learner perform the skill the course is named after, and grade it. Recorded audio, video, an image, or executed code is judged against the assignment rubric. A free chapter with nothing graded breaches the publish standard and is flagged rather than shipped quietly.
Grounded writers
Readings, worked problems, cheat material and quizzes are written against excerpts retrieved from the research corpus, and a claim about a tool, API or version has to come from one of those excerpts or be left out.
The ramp designer
Before any lesson is written, one pass authors the course arc: the pain the chapter opens on, the artifact carried through it, and which concepts are deliberately revisited later.
A curated reference library
Where the open web is thin, writers are grounded on license-verified reference works that either run or can be checked line by line. Each license is classified from the source own terms, and anything non-commercial or non-derivative is refused.
Citation locking
A citation has to resolve to a source the engine actually retrieved. Fabricated references are caught before publish instead of being discovered by a reader.
The simulated learner
A simulated learner attempts the chapter before a human does. A task that cannot be done with what the chapter taught, or a question it cannot answer, fails the gate.
Continuous re-tailoring
After each graded attempt, the rest of the course adjusts to what the learner actually demonstrated, and every adjustment is recorded.
AEGIS 1.5
Nothing ships that a gate has not read.
Deterministic fact-checking
Checkable claims are extracted and then verified by running or calculating them rather than by asking another model whether they look right.
Executed examples
Code examples are run in a sandbox during generation, and a failure escalates to a stronger model instead of falling back to prose that claims an output nobody ran.
Whole-chapter review
One long-context pass reads the entire free chapter and its quiz together, from a different model family than the writer, and orders targeted regenerations.
The skeptical-buyer read
A separate pass reads the finished chapter as someone deciding whether to pay, and its blockers are remediated before publish.
Publish standard and defect ledger
Every gate outcome is recorded per course. A block that cannot be repaired ships visibly marked, so a hole is never invisible.
AEGIS 1.0
Courses are researched, not recalled.
Research-grounded coverage
Before anything is written, the goal is researched and a coverage specification is built from current sources instead of from what a model remembers.
Objective-driven curricula
The course is generated against measurable objectives and a knowledge graph, so progress is defined by what the learner can do rather than by pages turned.
Coverage critic
A second pass re-reads the generated outline for gaps against the specification and regenerates what is missing.
Mastery write-back
Every assessment result updates the learner mastery, review interval and next review date, so the same course adapts as the learner changes.
This page describes mechanisms that are in the pipeline today, not a roadmap. The gates are measured on the free chapter and their outcomes are recorded per course, but they do not catch everything, and we would rather tell you that here than have you find it in a lesson. If you are evaluating lrnit for an organization and want to see the ledger for a real course, ask and we will show you one.