llm-wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From toolkit to maintained project
LLM Wiki combines a Python CLI/MCP toolkit with instructions for an external wiki-maintainer agent. The human curates sources and asks questions; the agent is assigned reading, interpretation, cross-reference maintenance and semantic writing. The package scaffolds projects and pages, retrieves retained text and diagnoses structure. It is a knowledge-maintenance system, not an enclosing agent runtime. The claim that knowledge is compiled once, kept current and compounds describes the intended workflow; it is not an established quality or benefit result. See RT-CLM-agent-bookkeeping and EPI-CLM-compiled-knowledge.
The evidence is SRC-1, https://github.com/cobusgreyling/llm-wiki at commit a424e6875aa2d5736d9c0d03d7c28bdbe32892a5, with cutoff 2026-10-14. The whole-system boundary covers shipped instructions, toolkit helpers and callers, not external host, model, qmd or editor implementations. Toolkit findings are code-grounded. Semantic maintenance is doctrine/design; populated examples and authored logs are illustrative state and reported operation, not authenticated execution traces. No target execution or causal experiment supports this account. See the boundary and evidence and Source register. This synthesis uses the supplied accepted members without further source reads.
wiki init copies bundled instructions, templates and seed files into a project and optionally initializes Git. The external host must load the maintainer instructions, supply model execution and permit filesystem editing. Their force is natural-language guidance, not toolkit-enforced permission. In particular, read-only raw sources and append-only operation history depend on compliance. Neither Git initialization nor the generated MCP configuration establishes automatic commits, rollback, prompt loading or a deployed isolation envelope. See RT-RTE-project-init, RT-CMP-agent and RT-BAP-maintainer-instructions.
Integrating a source and retaining the result
The representative multi-step workflow is incremental ingest of one human-selected raw document into an existing wiki. The maintainer is instructed to read it completely, view images separately and discuss takeaways before writing unless batch mode applies. It then creates a source summary, updates affected entity/concept pages, conditionally revises synthesis, flags conflicts, updates the catalog and appends the log. Delivery means this integrated file state, not merely a conversational report or source-page mapping. The same agent owns interpretation and dependent writes; no independent reviewer or transaction is established. See RT-RTE-source-ingest and the workflow account.
The retained layers serve different purposes. Raw originals are imported evidence inputs. Source pages retain interpretations and possible quotes. Topic pages combine incremental observations; synthesis holds a cross-source account; the contradiction ledger preserves opposed claims and resolution explanations. Filed answers retain prior analyses. The catalog guides discovery, while the operation log is authored history rather than a transcript of agent experience. Reasons can survive in quotes, evolution text and resolution notes, but later queries need not read them. See MEM-OBJ-raw, MEM-OBJ-sources, MEM-OBJ-topics, MEM-OBJ-synthesis, MEM-OBJ-contradictions and MEM-OBJ-log.
Human source selection and discussion do not establish individual claim approval. Optional answer filing likewise lacks an enforced acceptance handshake. The template writer is a separate route: it renders placeholders, rejects an existing destination unless forced, and leaves semantic integration and catalog repairs to its caller. Incomplete pages can therefore be available to retrieval. Interrupted multi-page work requires external inspection and repair; optional history is not automatic recovery. See RT-RTE-page-create and MEM-OBJ-answers.
Later questions and corrective work
For a question, the instructed agent reads the catalog, selects a search query, chooses relevant pages or headings, synthesizes an answer with citations and offers to file it. Built-in BM25 ranks current Markdown text, including frontmatter, and returns paths, scores and snippets. Expansion returns a requested page or heading block with descendants. These are implemented pull routes; model use and answer effects are unobserved. Result counts and section selection reduce particular tool responses, not total context residency. Full pages remain reachable; host compaction and isolated agent contexts are outside the implementation. See RT-RTE-wiki-query and its memory annotation.
Optional qmd search replaces the selector with an external subprocess and adapts its results. Its retained index, algorithm, provenance and refresh behavior remain uninspected. The adapter does not inherit every built-in selection or containment property. Metadata exploration returns pointers, links and counts rather than evidence passages. Recent-log parsing omits entry bodies and prescribed hyphenated operations such as answer-filed; full-log expansion remains an alternative. See RT-RTE-qmd-search, MEM-RTE-metadata-read and MEM-CLM-log-parser.
Diagnostics check structure more strongly than meaning. Source status calls a file ingested when raw_file associates it with a source page, even without a substantive summary. Watch reports newly pending names; its shipped callback does not ingest them or detect same-name content changes. Contradiction lint detects marker language and lexical ledger references, not incompatible propositions. CLI lint can fail an inspection at a severity threshold; MCP returns diagnostics without that exit policy. Neither is a semantic admission gate. No toolkit connection withdraws claims from search merely because they are stale or disputed. See EPI-RTE-source-coverage, RT-RTE-pending-watch, EPI-RTE-contradiction-check, RT-RTE-wiki-lint and MEM-ABS-withdrawal-gate.
What revision establishes
The strongest supported contribution is an inspectable place for persistent accounts, criticism and reasons, coupled to selective later reload. The protocol requests evidence-responsive revision rather than treating the first summary as final. That contribution does not establish improved capacity for future action: no comparison traces criticism of a consumed account through revision to a changed later decision. Source import has unknown warrant, and semantic input/output correspondence remains indeterminate. Citations, retention and cross-links do not supply a recorded evidence-consuming acceptance decision against a named criterion, intended use and scope. No produced accepted ampliative claim or subsequent lifecycle integration is established. See RT-RTE-source-ingest, MEM-BAP-wiki-content and the epistemic bounded conclusion.
The runtime's theory-builder assessment remains separate for each condition:
- 1, localized content: afforded. Pages can hold whole accounts; no particular theory is reconstructed here.
- 2, consumption: claimed. Instructions condition decisions on source implications; actual theory-dependent decisions remain unobserved.
- 3, content-directed criticism: uninspected. Conflict flagging does not show blame assignment and resulting revision or changed reliance.
- 4, iteration: uninspected. Persistence permits another round but does not demonstrate criticism shaping it.
These are the findings of RT-RTE-source-ingest, not a composite grade. Addressability is at page/heading level, not established assumption-level organization. Learning through criticism remains unestablished independently of the presence of editable memory. Reflection is also unestablished: method text represents conventions, but both causal directions between operation and representation are unobserved. Autonomous theory building is unestablished independently; human co-evolution and opaque model processing do not show computation performing every internal operation.
Instruction synchronization implements a distinct self-change affordance: changed canonical conventions are copied to fixed scaffold/example targets and can guide later agents if loaded. It neither automatically learns from logs nor migrates arbitrary deployed projects. At this project/session horizon, exercised self-improvement is unestablished: no independently specified improvement objective and evidence-shaped update determination are traced through changed organization, live instruction authority and subsequent behavior causally dependent on the change. Successful improvement is not measured either. This does not turn unknown trace-fed updates or provider adaptation into evidenced absences. See RT-RTE-instruction-sync, RT-CMP-agent and profile verification.
For repeated source integration and later question answering, the supported architecture offers reusable, editable interpretations and targeted reload; it does not guarantee current, faithful or accepted knowledge. A provenance-linked ingest trace with before/after pages and subsequent queries would change the operation assessment. Paired source/output checks and explicit acceptance criteria would change the warrant assessment. A trace connecting formulated criticism, blame, revision and a next-round decision would change the theory assessment. Evidence-shaped convention changes followed by causally dependent later operation would change the exercised self-improvement assessment. A semantic withdrawal gate, transactional maintenance or a verified qmd refresh contract would change the corresponding control boundaries. No ranking or comparative performance conclusion follows from the present evidence.
Limitations
The report verification declares no additional limits. There is no unresolved cross-report conflict. The two profile-verification limits are retained below alongside the evidence boundaries that prevent broader conclusions.
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No authenticated maintenance execution or causal comparison | SRC-1; RT-RTE-source-ingest, RT-RTE-wiki-query, MEM-OBJ-synthesis | Shipped protocol, helpers and sampled illustrative state at the frozen commit; target not executed | Successful delivery, faithful integration, recurrent use, improved future capacity or compounding benefit | Provenance-linked operations, retained inputs/outputs and later consumers; controlled comparisons for benefit attribution |
| External host and model internals excluded | SRC-1; RT-CMP-agent, RT-BAP-maintainer-instructions | Toolkit and host handoff configuration, not host implementation or provider | Prompt compliance, approval/isolation guarantees, model fixity or parameter adaptation | Host loading/permission evidence and provider version, parameter-change and endpoint contracts |
| Semantic transformation and acceptance unresolved | SRC-1; EPI-OBJ-wiki-pages, RT-RTE-source-ingest, RT-RTE-wiki-query | Instructions and retained page roles; no checked source/output correspondence or acceptance history | Entailed versus ampliative production, source warrant, per-claim approval and accepted-claim lifecycle integration | Checked correspondence plus recorded evidence-consuming acceptance with criterion, intended use and scope |
| Criticism, learning, reflection and exercised self-improvement not demonstrated | SRC-1; RT-RTE-source-ingest, MEM-OBJ-contradictions, RT-RTE-instruction-sync | Editable accounts, conflict ledger and convention propagation; opaque internal processing and unobserved later behavior | Conditions 3–4, improved future capacity, reflective/autonomous theory building and exercised evidence-responsive self-change | Claim-directed criticism with blame and retained next-round consequences; method/operation causal links; objective-linked update and subsequent dependent behavior |
| Optional qmd retained state is opaque | SRC-1; RT-RTE-qmd-search | Adapter and setup doctrine, not external index implementation or deployment | Complete substrate/form/lineage/authority inventory, guaranteed freshness after edits, backend selection equivalence | Index/update implementation, provenance and refresh contract tied to consumed project state |
| Parsed history is incomplete | SRC-1; MEM-RTE-log-read, MEM-CLM-log-parser | Heading regex and recent-log/stats consumers | Complete operation history or delivered reasons from parsed output | Parser covering all prescribed operation names and body delivery, or explicit full-log consumption |
| Template-consumer form is incompletely classified in the profile | SRC-1; RT-RTE-page-create; profile-verification limit | Profile distinguishes interpreted guidance but omits the renderer's separate fixed-rule consumption | Complete template-consumer form comparison; positive natural-language and symbolic findings remain valid | A separate classification unit for symbolic template substitution supported by the existing route record |
| Changed-template reload is grouped with unknown host instruction loading in the profile | SRC-1; RT-RTE-page-create, RT-BAP-maintainer-instructions; profile-verification limit | Known caller-requested software template read and uninspected external-host loading trigger | Complete route-level changed-template reload comparison; software reading must not be treated as equally opaque | Separate software pull and host-loading units; host trigger/selection evidence to resolve the latter |