Glossary
The terms IE uses, what is implemented, and what each term does not claim.
- AI Systems Architecture
- The system around a model: data inputs, context preparation, reasoning, validation, policy, human decisions, execution integration, feedback, and memory. The model is a component within that system, not its source of truth.
- Approved Project Memory
- Active project information that a human has explicitly approved. Pending, rejected, retired, and superseded records are not active project truth.
- Bounded Source Evidence
- Source snapshots, paths, hashes, or compiler inventory selected by an operator within an explicit scope. This evidence supports review but is not automatically approved truth.
- Closure Verifier
- A deterministic Batch 3 read model that reports missing links between planning history, result analysis, human review, execution feedback, pending proposals, and their approval or rejection.
- Codex Partner Mode
- The implemented manual AI-assisted workflow. IE prepares a task packet, the operator copies it into Codex, and the operator pastes the final report back for deterministic analysis and human review. IE does not call Codex directly.
- Codex Task Packet
- An implementation request prepared for manual use with Codex, including its scoped origin and evidence references. IE does not call a Codex API; the operator copies and pastes the packet.
- Context Pack
- A scoped handoff artifact assembled from project state, evidence, quality checks, risks, and the next action. A passing quality check does not certify truth.
- Contradiction Queue
- An implemented queue for deterministic contradiction findings and their human-review status. It does not detect semantic contradictions or change project truth automatically.
- Current-State Diagnosis
- A non-persistent, server-authoritative deterministic evaluation of available project state. Each finding is labeled as evidence, explicit inference, or unknown. The diagnosis is neither approved memory nor proof of correctness.
- Deterministic Control
- A check whose inputs and rules can be evaluated without model judgment. IE uses these controls to establish known facts, bound context, validate structure and references, and enforce workflow gates. They surround AI reasoning; they do not replace it.
- Execution Feedback
- The operator's record of a reviewed implementation outcome and any available build or test evidence, connected to the diagnosis, plan, and packet behind the work.
- Governed Lineage
- References connecting a diagnosis finding to the Brief, Outline, Plan, selected item, local ticket, packet, analysis, review, feedback, and memory proposal. These references show history; they do not grant approval.
- Guided Mode
- A seven-stage presentation of the implemented IE-on-IE workflow. It groups existing deterministic, manual, and human-reviewed actions without adding model calls, persistence, source editing, pull-request creation, or memory approval.
- Hallucination Gauge v0
- An implemented deterministic, read-only 0–100 operational risk score over available signals. It is not hallucination detection, an LLM judgment, or a truth score.
- Human Approval Gate
- An explicit operator decision required before a workflow artifact or memory proposal advances. Human approval remains the truth gate in the demonstrated workflow.
- IE Learn
- The first applied AI product inside the repository. Its implemented AI-102 path prepares a local evidence corpus, retrieves relevant chunks, generates grounded questions, and keeps question review, learner scoring, weak areas, readiness, and persistence outside the model.
- Grounded Question
- A generated question created from retrieved evidence and carrying references to the allowed source chunks used for generation. Grounding and deterministic checks support review; they do not make model output automatically correct.
- Weak Objective
- In IE Learn, an objective with at least three recorded attempts and accuracy below 70 percent. This is a deterministic practice signal, not an LLM judgment or an exam prediction.
- Memory Proposal
- A request to add project memory. Selected learning remains pending until a human separately approves or rejects it.
- Model Gateway
- The implemented IChatModelGateway boundary for chat and streamed completion. The current infrastructure includes an Ollama-backed implementation and local-model routing, but Guided Mode does not yet expose a live governed LLM review.
- Negative Memory
- Proposed, not implemented. A future memory type for rejected ideas, invalid assumptions, failed approaches, and blocked paths. Rejected proposals are not Negative Memory.
- Review Council
- Proposed, not implemented. A future pattern for multiple critic evaluations whose findings would remain evidence rather than truth.
- Roslyn Source Inventory
- An implemented read-only compiler-aware inventory with bounded previews and diagnostics. It is evidence, not a full dependency graph, call graph, semantic validation, or source editor.
- Session Integrity
- The target property that an agent workflow stays aligned with scoped evidence, approved memory, explicit constraints, reviewed state changes, and human intent. Supporting pieces and selected workflows are implemented; universal runtime enforcement is not.
- Restart Verification
- A deterministic, read-only reconstruction that compares the saved workflow after an API/client restart and reports source revalidation separately. It proves continuity, not correctness.
- Trusted Restart Pack
- An implemented v1/v2 format for export, validation, optional scoped source revalidation, and attaching untrusted material as evidence only. It supports restart review but cannot approve truth.