Skip to content
InvariantEngine

How it works

Seven stages showing what IE checks, what Codex proposes, and what a human decides.

The operating rule

Model output, including the report returned from Codex, is proposal material. Deterministic analysis gives the operator evidence, not proof. Humans confirm scope, accept planning, review the result, and separately approve or reject memory suggestions.

Guided Mode does not currently run a live governed LLM review. The AI-assisted step is a manual handoff: IE prepares context and instructions, the operator uses Codex, and IE receives the pasted report for validation and review.

Three different responsibilities

Deterministic

Establish source facts, prepare context, check references, analyze the pasted report, and verify workflow state.

AI-assisted

Codex reasons about the prepared task through a manual copy-and-paste exchange in the current demo.

Human-reviewed

The operator chooses the work, accepts the plan, reviews the result, and approves or rejects suggested memory.

  1. 01

    Ground the project

    Batch 1 Human setup + deterministic scan

    The operator supplies the repository root, reviews the include and exclude scope, and confirms it before IE captures evidence.

  2. 02

    Diagnose the current state

    Batch 1 Deterministic

    IE evaluates server-loaded source, memory, onboarding, contradiction, Context Pack, drift, and compiler signals. Missing evidence stays unknown.

  3. 03

    Choose one improvement

    Batch 1 Human decision

    A human reviews the evidence behind one finding, chooses whether to continue, and edits the scoped requirement.

  4. 04

    Build and review the plan

    Batch 2 Deterministic structure + human gates

    IE creates the Brief, Outline, and Plan as separate records. The operator reviews and accepts each one before the work can advance.

  5. 05

    Prepare the AI handoff

    Batch 2 Manual AI handoff

    IE turns one accepted plan item into a scoped task packet. The operator manually copies that packet into Codex; Guided Mode does not call a model here.

  6. 06

    Validate and review the result

    Batch 3 Deterministic check + human review

    The operator pastes the Codex report into IE. Deterministic analysis checks the report, then a human accepts it, accepts it with fixes, requests rework, or rejects it.

  7. 07

    Record feedback, learn, and verify

    Batches 3–4 Human decision + deterministic verification

    IE records execution feedback, keeps selected learning pending until a separate human decision, checks closure, and can verify the saved workflow after restart.

Synthetic workflow illustrations

These are illustrations, not live product screenshots. They summarize the implemented Batch 1–4 flow without exposing private repository or environment data. Labels are generic and local identifiers are omitted.

Synthetic Invariant Engine operator view with a confirmed bounded source scope and deterministic diagnosis.
SYNTHETICIllustration — the operator confirms the source scope, reviews the available evidence and unknowns, and selects one editable improvement.
Synthetic Invariant Engine planning view connecting a diagnosis finding to accepted Brief, Outline, Plan, ticket, and Codex packet.
SYNTHETICIllustration — the Brief, Outline, Plan, selected item, local ticket, and paste-ready packet stay connected through explicit references.
Synthetic Invariant Engine closure view with result analysis, explicit human review, execution feedback, pending memory, disposition, and closure verification.
SYNTHETICIllustration — result analysis, human review, and execution feedback come before a pending memory suggestion, a separate approval decision, and the closure check.
Synthetic Invariant Engine restart verification view with equivalent durable links, fresh bounded source evidence, and an explicit correctness boundary.
SYNTHETICIllustration — fresh API and client processes reload the saved workflow, compare its records, and show source revalidation warnings without claiming code correctness.

Current boundary

The full workflow has survived a real API/client restart against durable state. That proves the demonstrated private-alpha workflow can be reconstructed after restart. It does not prove code correctness, make AI output true, or enable autonomous improvement. IE is not generally production ready, and the repository remains private.

Batch 5 adds the clean-checkout repeatability rehearsal, operator runbook, and sanitized review package. It verifies that the demonstrated process can be repeated; it does not add model, execution, or Engine capability.