About InvariantEngine

I'm Joshua Couto. I started building Invariant Engine because I wanted AI to help with real engineering work without losing the context, evidence, and decisions that make the result understandable later.

The problem was not that models could not produce code or plans. The problem was operating them safely around a real project. If the model receives the wrong context, invents a reference, or turns an assumption into project memory, a fluent answer can still leave the system in a worse state.

I treat the model as a reasoning component, not the authority for project truth. Its output can be useful and still remain a proposal. IE keeps the source facts, context preparation, validation, human decisions, execution feedback, and approved memory outside that boundary.

Why the layers matter

Context determines what a model can reason about. Provenance lets me check where that context came from. Deterministic validation catches structural and policy failures without asking the same model to grade itself. Human review is deliberate because accepting a plan, an implementation result, or project memory changes what the system is allowed to do next.

IE Learn made the problem concrete

The learning application came before much of the general framework. I built an AI-102 workflow that prepares a corpus, retrieves evidence, generates grounded questions, and records practice results. That work forced the model boundary into focus: the model could generate a useful question, but application code still needed to own evidence references, quality checks, publication state, correctness, weak areas, and readiness.

IE Learn became the first applied AI product inside the repository and helped shape Invariant Engine around context, provenance, deterministic decisions, explicit review, and persistent state. The IE Learn page explains the implemented path and its limits.

Why I test IE against itself

I use IE against its own private repository because it gives me a real engineering workflow with consequences I can inspect. I confirm scope, choose a diagnosis finding, review the plan, manually exchange a task and report with Codex, review the result, and separately decide whether any learning belongs in project memory.

The codebase also contains model-gateway, local Ollama, reasoning-session, and streaming foundations. Guided Mode does not yet expose those pieces as a live governed LLM review. The public demo currently proves the context, control, handoff, review, feedback, and memory boundaries around AI-assisted work.

Private-alpha status

The product and source repository remain private. Batches 1–5 are implemented and repeatable from a clean checkout; Batch 5 covers repeatability and the sanitized review package rather than new Engine behavior. IE does not autonomously edit source, create pull requests, or call Codex, and I do not present model output as verified truth. It is a private alpha, not a generally production-ready system.

The full workflow has survived a real API/client restart against durable state. That test verifies continuity of the saved workflow. It does not verify source-code correctness or make AI output true.

Public access

There is no public repository link because the repository is private. For now, this site is the public record of what I have implemented and what remains outside the product boundary.