LLM Wiki v2 synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From supplied material to a durable knowledge bundle
LLM Wiki v2 is a Python-controlled knowledge pipeline, not an enclosing agent that chooses arbitrary tools. It files documents or client-supplied memories into a folder tree of Markdown notes, then answers separately requested questions from selected retained material. This account covers the shipped library, CLI, HTTP/web application, MCP integration, maintenance and evaluation consumers at SRC-1 commit e0a205afe60bd1913d083a3ed4cdab81e38e4c26, with cutoff 2026-11-08. The whole-system boundary includes classic and CROW modes, not the external MCP client's agent loop or provider internals. Evidence is inspected implementation, shipped instructions and attributed benchmark reports. No target execution or causal experiment establishes successful operation here (boundary, Source register).
The central division is between semantic proposals and code-owned progression. A caller supplies text and optional origin, title and fields. Preparation produces a candidate before reading the current tree. Locked filing rechecks origins, reloads state, routes the candidate, restricts merge candidates by conflicting fields, matches a note and decides whether to consolidate. Code then merges or creates, saves raw provenance and the note, and attempts related-link enrichment. Starting ingest does not approve each generated sentence: model/software decisions admit ordinary rewritten content without per-note human sign-off. Python checks types, names, tree membership and other structural preconditions, not every proposition's truth (RT-RTE-source-filing, EPI-RTE-step-check).
Classic asks a generative model for semantic decisions. CROW replaces route, match, consolidation, relation and retrieval decisions with classifier calls while retaining classic writing. CROW retrieval can fall back to classic on empty selection or classifier ModelError; missing/refused credentials instead fail configuration. Batch preparation uses threads, followed by ordered locked filing. Planning can place items directly, and automatic splitting can change taxonomy. These are fixed alternate paths, not delegated autonomous workers (runtime account, RT-RTE-batch-sync).
The persistence boundary matters to recovery. Cooperating thread/process locks and stale-file checks protect writes, with atomic replacement of individual files. They do not provide a transaction across raw copies, notes, indexes, archives, links and logs, or an atomic compare-and-swap against arbitrary editors. A relating ModelError after admission can leave the note successfully filed without related links. Other failures may leave partial effects; forgetting cached tree state is not rollback. The representative workflow therefore supports caller-owned retry and local fallbacks, not universal automatic recovery (workflow account, RT-RTE-source-filing).
Retention preserves more than answering receives
The durable bundle separates extracted raw source text from rewritten note bodies and machine-consumed metadata. Folder descriptions, titles and summaries influence later placement and relevance. Hashes, fields, source references and predecessor paths control repeat-work suppression, merge/replacement eligibility and historical loading. Retained classifier responses in SQLite can substitute for new inference; benchmark replay separately retains writer and classifier responses under different keys. Neither cache is model-weight storage or evidence of current-provider behavior (MEM-OBJ-raw, MEM-OBJ-note-body, MEM-OBJ-note-metadata, MEM-OBJ-folder-index, MEM-RTE-answer-reuse).
Context is reconstructed for individual calls, not accumulated as a continuing conversation. Oversized sources are split into blocks; bounded part notes, rather than the original blocks, reach the combine call. Only the final note and raw source persist through this route. Missing-value detection at combination warns rather than vetoes. Raw retention therefore does not establish that omitted material survives into a later answer. Each call resembles an isolated input frame with a returned value, but stored paths are not automatically dereferenced: code must explicitly reload their content (MEM-RTE-source-reduction, memory read-back).
A later question starts a fresh Researcher. Selection uses structure and summaries before chosen note displays reach the answer writer. That display contains title, summary and main body, not most frontmatter or See-also links. Optional historical loading follows predecessor paths; spreadsheet expansion can add rows and, as report verification found, numeric-column statistics. The statistics are derived context, not necessarily values copied from selected rows. No selection yields a fixed no-answer message. Otherwise the result includes text and citation metadata; it is not automatically saved as another note (RT-RTE-question-answer, MEM-OBJ-note-body, report-verification limits).
Question/recall and explicit MCP note reads are requested retrieval. Internal filing also automatically supplies retained structure and matched content to decision calls; split and sync operations supply retained control state. These internal deliveries do not establish proactive memory injection into an enclosing client. Nor are all direct reads the same: MCP returns the selected display, whereas HTTP /note exposes body and frontmatter. The display omission is channel-specific, not global unavailability (MEM-RTE-note-read, MEM-BAP-retained-knowledge, profile verification, report-verification limits).
Consequential guards have local warrant
The system does more than store unchecked model output. Merge guards detect selected lexical value losses, request bounded retries and choose a retry when it loses fewer old values. Revision replacement checks generated-body untouchedness and source identity; classifier-equipped revision additionally filters selected old sentences against source passages. Failure of replacement eligibility falls back to merge. These checks can change what is filed, but value presence does not establish correct attribution, negation or factual completeness. Sentence agreement is relative to selected supplied source text, not external truth (EPI-RTE-merge-guard, EPI-RTE-guard-choice, EPI-RTE-revision-check, EPI-RTE-revision-admit).
At answer time, conflict classification over selected passages can add a request to report competing values and provenance. It neither adjudicates truth nor rejects the final answer, and runs before spreadsheet-row expansion. Parsed citation metadata is restricted to delivered note identities, but answer text is not rewritten or checked for entailment. The where fields filter is not authorization: historical expansion follows it, and direct reads are separate paths. Source-only answers and complete factual preservation remain prompt goals, not enforced guarantees (EPI-RTE-conflict-check, EPI-RTE-conflict-hint, RT-RTE-question-answer, RT-CLM-factual-preservation, EPI-CLM-source-answers).
Maintenance also has distinct consequences. Marking a source removed leaves derived notes available; rebuild can refile that content without its removal mark. Purge instead deletes notes citing the origin, including mixed-source notes, then refiles surviving sources. This is reference-based withdrawal, not semantic erasure of every copied claim or retraction from external clients. Archive replacement retains a predecessor for conditional historical reads. Route audit can propose and, with apply requested, perform moves; health and lint return diagnostics rather than factual endorsement (memory maintenance annotation, EPI-RTE-origin-purge, MEM-OBJ-archive, EPI-RTE-route-audit, EPI-RTE-audit-apply).
Adaptation, method criticism and assessed use
A concrete retained-experience contribution is wired: split attempts enter a durable log, and later splitting reads folder-specific attempt counts to suppress another attempt at the same count multiple. This is a local automatic trace-fed behavior-shaping update from operational events. It does not demonstrate improved future competence. Client remember can retain decisions and reasons, but the server does not establish their original trace inputs or automatic client extraction. Cache/replay, sync and custom update alternatives remain incompletely classified; the split-history witness neither resolves them nor supports a system-wide learning negative (MEM-OBJ-log, RT-RTE-batch-sync, profile verification).
The separate method-level questions remain bounded. Localized method instructions exist, satisfying the formulation minimum of theory-builder condition 1. Their delivery and operational wiring are established, but condition 2's dependence on what a formulated theory says is not demonstrated by delivery alone. Content-directed criticism of the system's methods and retained iteration on that criticism—conditions 3 and 4—are unestablished in the traced defaults. The shipped multi-pass example does critique and rewrite a note relative to its source; it must not be erased by a system-wide claim that no critic/rewrite route exists. It does not establish revision of the system's own methods. Improvement of future capacity from criticism of a consumed theory, reflective self-representation with both causal directions, and role-complete autonomous theory building are not established by these records (RT-OBJ-prompts, RT-BAP-model-instructions, epistemic bounded conclusion, report-verification limits).
Exercised self-improvement at this repository boundary is likewise unestablished: retained control changes can reach later decisions, but no supplied subsequent-operation trace closes causal dependence on an improvement-directed organizational change. The split-history path remains a positive architectural adaptation, not evidence of successful improvement. Prompt overrides, settings and custom Steps enable caller-directed changes, not a demonstrated autonomous method-revision cycle (RT-RTE-runtime-configuration, MEM-OBJ-log, epistemic bounded conclusion).
For a scenario of filing documents and later answering questions, the source establishes the necessary durable handoff, selective context construction, local checks and partial-success handling. It does not establish lossless transformation or reliably grounded answers. For experience retention through MCP, storage and later requested delivery are supported, while client extraction, isolation and behavioral activation remain external or unobserved. Client-model relay lets requests leave the server and replies resume filing; instructions recommending restricted subagents do not enforce client tool isolation (workflow account, RT-RTE-client-relay).
Paired source/candidate/final-note/answer traces would permit bounded preservation and grounding judgments. Retained criticisms linked to method updates and their later consumers would change the method-learning assessment; subsequent causally dependent operation would be needed for exercised self-improvement. Controlled comparisons varying one mechanism would support finer attribution than the repository's reported bundled comparisons. The inspected retrieval evaluator checks selected-note identity without answering; lexical fact scores test string presence, not proposition truth. Reported benchmark tradeoffs therefore do not settle answer correctness, successful learning or individual component effects (MEM-OBJ-evaluation-results, EPI-RTE-fact-score, EPI-CLM-benchmark-results).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Unresolved check-result part separation, carried from reconciliation and report verification | SRC-1; EPI-OBJ-check-results, MEM-OBJ-evaluation-results, report-verification limits | Distinct check routes are recorded, but individual result objects remain coarse containers | Complete part-level inventory or a common warrant, retention horizon or authority inferred from the containers | Separate accepted object records connecting each result part to its producer, persistence and consumer |
| Direct-read inventory is incomplete in analyst records | SRC-1; MEM-RTE-note-read, MEM-OBJ-note-metadata, RT-RTE-question-answer, report-verification limits | MCP display and answering traced; verifier additionally inspected HTTP /note in src/llmw2/server.py |
Generalizing display omissions to every consumer, or claiming metadata/related-link text is globally unavailable | Accepted route/object coverage for HTTP body/frontmatter delivery and its consumers |
| Spreadsheet expansion includes incompletely recorded arithmetic | SRC-1; MEM-OBJ-raw, RT-RTE-question-answer, EPI-OBJ-answer, report-verification limits | Verifier traced src/llmw2/bundle/sheets.py numeric count/sum/min/max over all rows when over row budget |
Complete derivation inventory or assuming all numerical answer context is copied from selected rows; source truth remains separate | Accepted statistics-route records including parsing, domain, input coverage and delivered context; paired outputs for semantic claims |
| Selected defaults are not every shipped custom workflow | SRC-1; workflow account, RT-RTE-runtime-configuration, epistemic bounded conclusion, report-verification limits | Verifier inspected examples/multi_pass_summarize.py source-relative critique/rewrite and examples/custom_select.py deterministic selection/metadata answer; neither executed |
Exhaustive workflow inventory or system-wide absence of critic/rewrite and non-model answer routes | Accepted route records for those examples and any further material custom paths; execution evidence for exercised properties |
| No linked execution or causal comparison closes activation and benefit claims | SRC-1; RT-RTE-source-filing, RT-RTE-question-answer, MEM-OBJ-log, EPI-CLM-benchmark-results | Implementation, prompts and attributed aggregates; no independently inspected paired executions | Demonstrated preservation, grounded answers, improved future capacity, exercised self-improvement or component-level causal benefit | Paired end-to-end traces, retained-update/later-consumer traces and appropriately controlled comparisons |
| External principals, providers and injectable implementations remain unresolved | SRC-1; RT-CMP-llm, RT-CMP-classifier, RT-RTE-client-relay, RT-RTE-manual-maintenance, MEM-OBJ-answer-cache, MEM-OBJ-sync-state | Repository integration code; not external client/provider/connector internals or arbitrary injected code | Immutable provider identity, human content approval, client isolation, complete custom-memory classifications or automatic host experiential learning | Pinned implementations and authenticated provenance plus bounded client/provider traces, under an expanded evidence boundary |
| Evaluator coverage and reproducibility are partial | SRC-1; MEM-OBJ-evaluation-results, MEM-OBJ-replay, EPI-RTE-fact-score, EPI-CLM-benchmark-results | Retrieval/replay/lexical scoring inspected; other families incompletely traced, contradiction/revision harness absent and private corpora unfrozen | Complete evaluator taxonomy, independent reconstruction of reported results, propositional accuracy or current-provider/cap behavior inferred from replay | Available harness/corpora and retained executions; complete evaluator tracing and metrics appropriate to the claimed semantic outcomes |
No source rereads were needed for this synthesis; all findings use the supplied members.