Capabilities
What exists in the platform, what the current demo uses, and what remains deferred.
MVP rebaseline: July 29, 2026 · Guided Mode presentation: July 30, 2026.
What the current Guided Mode demo uses
The demonstrated path uses scoped repository evidence, deterministic diagnosis, reviewed planning, a manual Codex handoff, deterministic pasted-result analysis, human review, execution feedback, separately approved or rejected memory suggestions, and restart verification.
IE also has model-gateway, local Ollama, reasoning-session, and basic streaming foundations. Guided Mode does not currently use them for a live governed LLM review.
APPLIED PRODUCT / IMPLEMENTED AI-102 PATH
IE Learn uses the AI foundations directly
IE Learn has a local content pipeline, deterministic chunking and provenance, Ollama embeddings, in-memory vector retrieval over JSONL datasets, grounded question generation, question quality and publication states, practice sessions, weak-area targeting, readiness calculations, APIs, a Blazor UI, and configurable EF persistence.
It consumes prepared transcript text but does not transcribe media. It does not crawl GitHub, use a managed vector database, or implement AZ-104/AZ-305 study paths.
Review the IE Learn pipeline and boundaries ->AI and reasoning foundations
Implemented foundationThese platform pieces exist in the codebase. They are not presented as a live AI review inside Guided Mode.
Model gateway and local routing
An IChatModelGateway boundary supports chat and streamed completion. The current infrastructure includes an Ollama-backed implementation and configurable local-model routing.
Reasoning sessions and basic streaming
Reasoning sessions can assemble project context, stream model chunks, publish basic SignalR notifications, and retain session exchanges through the event/read-model foundation.
Context and evidence
Implemented todayThese capabilities establish what the system knows before any model is asked to reason about it.
Scoped repository evidence
The operator confirms include and exclude rules before IE captures bounded source snapshots, hashes, and provenance.
Read-only Roslyn inventory
IE can load a compiler-aware inventory with bounded previews and diagnostics. It is evidence, not a dependency graph, semantic proof, or source editor.
Context Packs
IE assembles approved memory, source evidence, workflow state, risks, quality findings, and a next action into Markdown or JSON for review and handoff.
Approved project memory in context
Active, human-approved project memory can be loaded into session context. Pending, rejected, retired, and superseded records are kept out of active truth.
Governance and validation
Implemented todayThese checks establish facts, enforce workflow policy, and keep evidence or model output separate from approved truth.
Deterministic current-state diagnosis
Server-loaded evidence produces strengths, gaps, blockers, warnings, unknowns, and a next action without an LLM call.
Quality, drift, and reference checks
IE checks source freshness, required references, workflow completeness, and restart equivalence independently of model judgment.
Contradiction Queue
Deterministic contradiction findings can be saved and reviewed with durable state. Semantic contradiction detection and automatic truth changes are not implemented.
Hallucination Gauge v0
A read-only deterministic risk score summarizes available workflow signals. It is not hallucination detection, model judgment, or a truth score.
Human gates and closure verification
Planning acceptance, result review, and memory approval remain human decisions. A separate verifier reports missing workflow links.
Execution and learning
Implemented demo workflowThe current IE-on-IE demo proves the handoff, review, feedback, and learning path without automating Codex or source changes.
Reviewed planning history
The selected diagnosis finding stays connected to the Brief, Outline, Plan, selected item, and local implementation task.
Paste-driven Codex Partner Mode
The operator copies a prepared task packet into Codex and pastes the final report back into IE. There is no Codex API integration.
Result analysis, review, and feedback
IE analyzes the pasted report with deterministic heuristics, then records the human review, execution feedback, and available build or test evidence.
Pending learning with separate approval
Selected learning can become a pending memory suggestion only. A human must separately approve or reject it before it can affect active project memory.
Restart and repeatability verification
The workflow has survived a real API/client restart against durable state. Batch 5 adds a clean-checkout rehearsal, runbook, and sanitized review package.
Deferred or not implemented
These are roadmap or architecture topics, not current product claims.
- Live governed LLM review inside the current Guided Mode workflow
- Codex API integration or autonomous implementation execution
- Autonomous source editing or pull-request creation
- Semantic contradiction detection and automatic packet invalidation
- Negative Memory, Review Councils, and Context Compression
- A production Hallucination Gauge beyond deterministic v0 risk scoring
- General production readiness and public repository access
Current limitations
- Output from models, including Codex, is proposal material, not verified truth.
- Guided Mode does not currently expose a live governed LLM review; its demonstrated Codex loop is manual and paste-driven.
- Invariant Engine does not autonomously edit source code.
- Invariant Engine does not autonomously create pull requests.
- There is no Codex API integration or autonomous implementation execution in the demonstrated workflow.
- Evidence and deterministic checks support review but do not prove correctness.
- Memory suggestions remain pending until a human separately approves or rejects them.
- Restart verification proves continuity of saved state, not source correctness, model-output truth, or production readiness.
- Human approval remains the truth gate.
- The source repository remains private.
Private-alpha status
The product and source repository are private. This site documents a scoped workflow operated by a human; it is not a source release, and the product is not generally production ready.