iLiC Notes 009
Evidence-Grounded Conversation
A public engineering note on separating fluent conversation from usable knowledge by routing answers through indexed evidence, citations, review artifacts, and governance boundaries.
Core Claim
iLiC is beginning to separate fluent conversation from usable knowledge. A response should not be treated as reliable because it sounds confident. A response becomes usable when it can be traced to indexed source evidence, cited, evaluated, and kept inside governance boundaries.
The Current Loop
The working loop is:
User question → Conversation planning → Source-library indexing → Source retrieval → Chunk inspection → Evidence-grounded answer → Reasoning trace → Claim Ledger record → Human review before memory or KB promotion.
What Changed
The newest iLiC layer introduces document-grounded answering from the Personal Internet Library. When the user asks a question, iLiC can classify the question, avoid stale local documents for current-events needs, search local source chunks, inspect matching text, compose an answer from retrieved evidence, cite source chunks with internal links, report which files and chunks were examined, and explain why the answer is the most supported working answer.
The important boundary remains: this process does not automatically write memory or promote knowledge.
Why This Matters
Most chat systems collapse conversation, memory, knowledge, confidence, inference, and source use into one output. iLiC is trying to keep those layers separate.
Conversation is not knowledge. Memory is not knowledge. Source material is not knowledge. Validated evidence can support a working answer. Governance decides what becomes durable.
Current Architecture
The Conversation Planner breaks stacked human questions into smaller units and labels them. The Evidence Answerer searches the Personal Internet Library and composes an answer from retrieved chunks. The Claim Ledger stores the review artifact: what was claimed, what evidence was used, what source files were involved, what confidence was assigned, and whether the result is supported or unverified.
None of these layers replaces the FSM. The planner prepares the route, the answerer drafts from evidence, and the ledger records the claim. Governance still controls what becomes durable.
Demonstrated Behavior
The current test pattern shows that iLiC can answer from more than one local document. One document can contain part of a subject, another document can contain another part, and a single related question can cause iLiC to retrieve both, cite both, explain the evidence path, and leave memory unchanged unless human approval happens.
This is a meaningful step beyond hard-coded responses. It shows a path toward answers that can be inspected after they are given.
What This Is Not Yet
This is not broad expert reasoning. It is not a claim that every answer is correct. It is not autonomous learning. It is not public-ready proof of product quality. It is a controlled step toward auditable conversation, where the system can explain what it used, why it used it, and what boundary still applies.
Next Proof Needed
The next clean owner demo should add real owner-authored documents to the Personal Internet Library, ask Wizard a natural question related to those documents, confirm iLiC plans the question, confirm relevant chunks are retrieved, confirm the answer includes citations, confirm Wizard shows the background audit trail, and confirm memory remains unchanged unless human approval happens.
Boundary
Retrieved source evidence may support an answer. Human approval is required before memory or KB promotion. The FSM remains the decision authority. The LLM remains a text-generation subroutine only.
What Is Intentionally Not Disclosed
This note does not disclose private source-library contents, memory contents, prompts, credentials, internal validators, or implementation details that would weaken the security of the system. It publishes the public design thesis: useful AI conversation should be able to show its evidence path.
References
- Prior iLiC Notes on governed cognition, memory, Truth Ledger, and governed learning.
- Local-first knowledge management practices around owner-controlled source libraries and citation trails.
- AI evaluation practices around retrieval, provenance, and evidence-grounded response generation.