An LLM Wiki gives an agent durable memory. A UAI-style assertion and handoff boundary gives that memory a way to reject claims that violate hard constraints. Used together, they form a practical pattern for project handoffs where context must compound without letting mistakes become permanent truth.
Non-normative report
This article is an explanatory LlmWikis report. UAIX.org remains canonical for UAI-1, Project Handoff, schemas, validator behavior, roadmap, and governance. This page does not claim certification, endorsement, or live .uai generation support.
The core problem
Traditional retrieval-augmented generation is useful for search, but its synthesis is usually temporary. A model retrieves chunks, answers a question, and then the structure of that answer disappears. An LLM Wiki moves synthesis into a persistent artifact: raw sources remain read-only, the compiled wiki records concepts and contradictions, and a compact schema file tells the agent how to maintain the graph.
That persistence is powerful, but it creates a new risk. If an agent misreads a source during ingestion, the error can be written into the wiki and reused by later sessions. For engineering, compliance, safety, or project-handoff work, a fluent summary is not enough. Some claims must be checked against deterministic rules before they are allowed to become memory.
Where UAI fits
A UAI-style boundary sits between probabilistic language generation and the harder world of project constraints, schemas, external validators, and approval workflows. The language model can propose an update. The assertion layer decides whether the update is allowed, blocked, or escalated into a structured handoff.
| Layer | Job | Failure it reduces |
|---|---|---|
| Raw sources | Keep original PDFs, transcripts, specs, tickets, datasets, and decisions unchanged. | Loss of evidence and accidental source rewriting. |
| LLM Wiki | Compile source material into linked concepts, entities, summaries, contradictions, and logs. | Repeated re-reading and context loss across sessions. |
| UAI-style assertion boundary | Validate candidate claims, artifacts, or handoff records against structured rules before persistence. | Hallucinated or constraint-breaking claims becoming durable facts. |
| Project handoff | When rules conflict, stop autonomous progress and return a structured choice to a human. | Silent relaxation of constraints or invented workarounds. |
Validated ingestion pipeline
The agent reads a new source and extracts atomic claims, decisions, constraints, and open questions.
Candidate updates are tested against schema rules, validator outputs, source requirements, or project handoff constraints.
Passing claims update the wiki. Failing claims become warnings, contradiction records, or handoff items instead of silent facts.
The wiki records source, claim, result, reviewer state, and any human decision that changed the operating boundary.
Project handoff under deadlock
The most important behavior is not automatic success. It is a disciplined stop. If a project has incompatible requirements, the agent should not invent a compromise or quietly relax a rule. It should preserve the conflict and hand it back as a decision package.
| Report field | Purpose | Example content |
|---|---|---|
| System status | Names the halt condition. | Blocked: constraints are irreconcilable under current rules. |
| Conflicting rules | Shows exactly which requirements collide. | Rule A requires one bound; Rule B requires an incompatible bound. |
| Evidence trail | Links source documents, validator traces, and wiki pages. | Raw source path, compiled concept page, log entry, validator output. |
| Resolution menu | Returns explicit human choices with tradeoffs. | Preserve all constraints, relax one requirement, change scope, or add a new resource. |
| Decision record | Stores who approved the change and why. | Approved option, date, rationale, follow-up tasks, affected pages. |
How the wiki compounds after handoff
Once the human decision is made, the LLM Wiki becomes more useful than a ticket comment or chat transcript. It can create a durable page for the deadlock, update related concept pages, add the decision to the log, and link the approved rule change to future work. The next agent does not need to re-discover why the compromise exists. It can read the evidence trail, follow the links, and respect the current boundary.
This is the practical value of combining the two patterns: the wiki remembers, and the assertion boundary prevents memory from drifting away from reviewed constraints.
Recommended operating pattern
- Separate source and synthesis. Keep raw sources immutable and compiled pages editable.
- Require claim status. Mark claims as source-backed, draft, contradicted, stale, or human-approved.
- Validate before persistence. Do not let constraint-sensitive claims become wiki facts until the relevant check has passed or a warning is attached.
- Escalate real conflicts. Use a structured handoff record when the agent cannot satisfy all rules at once.
- Record the resolution. Link the decision, affected pages, source evidence, and reviewer state so future sessions inherit the right context.
Related sources
- UAIX UAI-1 specificationCanonical UAI-1 source
- UAIX Project HandoffCanonical public handoff pattern
- RAG vs Incremental CompilationLlmWikis architecture comparison
- index.md and log.mdDeterministic navigation and audit memory
- Content Quality AuditEditorial quality implementation