llm-wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From sources to a retained wiki
llm-wiki is a knowledge and context-engineering system, not an enclosing agent runtime. It ships instructions for acquiring sources, compiling interconnected Markdown articles, answering questions and maintaining retained material. External Claude Code, Codex, OpenCode or Pi hosts supply inference, scheduling, tools and permissions. Python helpers implement local checks, session handling and selected admission controls. This account covers the repository's shipped responsibilities at commit 95a042c68998559305bc69ac8c482c4e1a415ad3, with an applicability cutoff of 2026-10-17. SRC-1 supplies implementation and doctrine/design evidence only. No target, tests, examples, providers or services were executed; no observed-run or causal-experiment evidence is registered. See the boundary and Source register.
The ordinary progression is source acquisition → typed raw files → model-planned article synthesis → links and indexes → later query. The foreground model resolves the topic, selects incremental sources using ingestion dates, reads them and writes or revises articles. A persisted topic guide can constrain article boundaries and scope: advisory deviations are reported; strict deviations request confirmation before writing. Ordinary article admission does not require per-article human approval. Self-checks follow writes and do not form an atomic rollback transaction. Articles can also originate in conversation rather than raw sources. See MEM-RTE-admit-sources, RT-RTE-compile-wiki and MEM-OBJ-schema.
Retained source and article bodies supply interpreted evidence; frontmatter and indexes supply dates, paths, summaries and selection cues. These parts have different consequences. A source link preserves a pointer, not a complete derivation. A verified date cannot certify human confirmation because compilation itself requests today's date. Count-based index freshness cannot detect every changed summary or same-count edit. Raw preservation is also conditional: collection ingestion requests new versions, whereas refresh explicitly permits raw overwrite. Conversation authorship and overwrite therefore qualify the claim that all knowledge is rebuildable from raw sources. See MEM-OBJ-raw, MEM-OBJ-wiki, MEM-OBJ-frontmatter, MEM-OBJ-index, MEM-RTE-refresh-sources and RT-CLM-concurrent-writes.
Research delegation and delivery
The representative workflow is standard, single-round plan-mode research into an existing active wiki, with personal specialists excluded. It is a connected shipped procedure, not a reconstruction from unrelated capabilities. The foreground model proposes three to five paths and waits for user confirmation. Each path-agent receives selected context and the full search → credibility → ingest assignment, then delegates search to specialized sub-agents. Each path-agent reviews and ranks its returned sources before writing its selected raw files. Foreground compilation happens only after path returns; it is not a pre-ingest approval gate. Source credibility criteria can reject candidates, but source selection does not establish the truth of each admitted proposition. See the workflow account, RT-RTE-research-paths, EPI-RTE-source-review and EPI-RTE-source-selection.
Path-prefixed raw names and deferred index updates coordinate parallel ingestion; one foreground compiler then reads the newly retained sources across paths. These instructions provide no filesystem ownership enforcement or protection from another session editing an article. Same-file edits otherwise accept last-write-wins. Retained plan status supports resumption: completed paths are skipped, interrupted paths become pending, and failed paths offer retry or skip. Partial raw files remain for deduplication rather than rollback. The exact skip-state transition at the compilation barrier is unspecified. Delivery is instructed article/link/index persistence plus logs and a report of sources, results and remaining gaps, not separate semantic acceptance by a receiver. All these semantic and coordination steps remain doctrine, with operation uninspected. See RT-RTE-research-paths and the workflow account.
Later consumption and operational memory
Compact querying pulls a selected topic's indexes, exact files and, if necessary, bounded search results into an answering model. It requests a cited answer or an explicit gap. Index-first selection and minimum-file guidance limit reads but are not token budgets. Query-lite and the explicit query command forbid writes; ambient full-skill indexing instructions instead request repair and logging. Pi launchers request read/search tools and disable ambient discovery, but forward caller arguments and depend on external host parsing and permissions. This is a route-specific read-only policy, not a whole-system sandbox. See RT-RTE-query-wiki, RT-OBJ-query-profile and MEM-OBJ-index.
Session memory has a stronger executable chain: hooks normalize and retain bounded event/state metadata, render templated digests, and can supply later context. At configured start/prompt events, code selects cwd-matched states with digest paths, orders them by recency and defaults to three. It pushes state-derived summaries and digest pointers, not digest bodies. Explicit rehydrate/show are separate requested reads; explicit digest promotion reads the body into topic raw notes. Host compaction and child-context disposal remain external. Missing helper configuration defaults to enabled/balanced despite opt-in wording in the skill. Manual summaries bypass the payload-redaction function, so the digest label establishes no blanket privacy guarantee. See RT-RTE-session-bridge, its memory annotations, MEM-OBJ-session-state and MEM-OBJ-session-digest.
Feedback capture separately classifies user text with regexes and templates, retains candidates and permits selected promotion into raw notes. It can reject simple acknowledgements, but requires neither the challenged earlier result nor a diagnosed root cause. Promotion is operational admission, not factual acceptance or automatic instruction installation. Conversation lesson extraction is a different, model-directed procedure: it requests context, symptom, cause, fix and Rule, permits article appends, and requires individual approval for proposed agent-instruction edits. See MEM-RTE-feedback-candidates, EPI-RTE-feedback-classification, EPI-RTE-feedback-promotion and MEM-RTE-extract-lessons.
What checks can establish
The system does prescribe substantive semantic checking. Thesis research scopes a claim, seeks opposing evidence and retains verdicts, caveats and change criteria. Audit can pursue supportive and adversarial evidence; librarian findings support maintenance. These are model-interpreted protocols, not executable factual adjudicators. Audit reports do not automatically exclude contradicted articles from later answers. No supplied history establishes an actually accepted ampliative claim, checked factual entailment or post-acceptance integration; article retention alone supplies none of these. See EPI-RTE-thesis-verdict, EPI-RTE-trust-review and EPI-CLM-compiled-knowledge.
Executable checks have narrower targets. Lint checks structure and references, not factual truth. Adapter dispatch checks registration, plan hashes and receipts but trusts adapter-reported remote verification; its subprocess is not an OS sandbox. Checkpoint seal/verify checks one fixed bundle's hashes, scanner findings and word-retention floor; semantic minimization and final publication remain separate decisions. Specialist enablement checks structure and membership, not method quality or new tool authority. Fixture grading can check terms, length and structure without understanding an answer's meaning, and does not automatically deploy a winning profile. See RT-RTE-lint-wiki, RT-RTE-adapter-execute, RT-RTE-checkpoint-export, RT-RTE-specialist-enable and EPI-RTE-fixture-evaluation.
Learning and change: separate conclusions
The strongest supported contribution is an implemented trace-fed path to durable operational summaries/candidates and later advisory context, plus prescribed lesson extraction and persistent editable guidance. This supports the memory comparison's local trace-learning finding, which concerns automatic durable trace-fed writes, not demonstrated improved capacity. Whether criticism of a consumed theory improved future action remains uninspected: no comparison connects a particular criticism, revision and later capacity. See MEM-RTE-feedback-candidates, MEM-RTE-extract-lessons, RT-BAP-session-context and profile verification.
Theory-builder conditions also remain separate. Localized topic guides and specialist methods afford condition 1; their prescribed application supports claimed condition 2; actual content-directed criticism and iterative reuse, conditions 3 and 4, remain uninspected on those routes. Thesis doctrine additionally claims opposing-evidence criticism and next-round reuse, but supplies no exercised instance. Lesson extraction claims diagnosis while leaving actual criticism and revised-rule iteration unobserved. These findings do not establish a theory builder meeting all four conditions in operation. See RT-RTE-compile-wiki, RT-RTE-specialist-enable, EPI-RTE-thesis-verdict and MEM-RTE-extract-lessons.
Reflection-oriented protocols represent research gaps or session experience and request changes to later work; demonstrated reflection's two causal directions are not established. Autonomous theory building is independently uninspected: model ownership of requested steps is not evidence that computation performed every internal operation. Exercised self-improvement of the bounded knowledge system is likewise unestablished. Retained guidance and rule proposals afford behavior-shaping change, but no supplied trace closes evidence → live organizational update → later operation causally dependent on that update. Successful improvement would additionally require outcome evidence. External model parameter changes are uninspected and no immutable weight version is pinned. See EPI-RTE-thesis-verdict, MEM-RTE-extract-lessons, RT-RTE-specialist-enable and RT-CMP-host-model.
For the selected scenarios, the supported assessment is therefore differentiated: research offers a model-mediated delegation/admission/delivery protocol; compact query offers bounded evidence-selection instructions; session continuation has executable selection and supply; export and private operations have bounded protocol gates. None establishes successful execution, useful recall, reliable truth judgments or measured savings. Provenance-linked research and recovery traces would test workflow operation. Source-to-article histories with explicit evidence-consuming decisions would change the semantic assessment. A retained criticism-to-method revision followed by dependent later use and a controlled outcome comparison would change improvement attribution. Host enforcement evidence, independent remote verification or stronger semantic export checks would change the corresponding safety claims—not retrospectively turn existing policy into an invariant. See report verification and the route records above.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operation or causal experiment | SRC-1; RT-RTE-research-paths, RT-RTE-query-evaluation, RT-RTE-session-bridge | Frozen source implementation and instructions, not real user stores or executions | Actual adherence, successful recovery, useful recall, accuracy, savings and causal benefit | Registered provenance-linked traces and appropriately controlled outcome comparisons |
| Incomplete route and payload inventory, carried from report verification | SRC-1; MEM-OBJ-datasets, MEM-OBJ-inventory, MEM-OBJ-projects, epistemic coverage, report limits | Selected compile/query/session/maintenance paths; native dataset transformations, collection fetch, output admission, package transformations and evaluator consumers not exhaustively traced | Exhaustive route counts, cross-host equivalence, complete payload classifications and system-complete epistemic negatives | Accepted records tracing those remaining producers, transformations, admissions and later consumers; opaque payload evidence where needed |
| Refresh delegation conflict remains unresolved | SRC-1; MEM-RTE-refresh-sources, EPI-RTE-source-refresh, report limits | Same refresh command: memory leaves delegation uninspected; epistemic calls it inapplicable to the foreground protocol; Agent permission specifies no packet | Settled refresh context boundaries or a complete isolation comparison; sampled comparison, user choice and overwrite remain supported | Explicit delegated packet/return and owner contract, or evidence resolving whether delegation belongs to this route |
| Failed-path skip transition is unspecified | SRC-1; RT-RTE-research-paths, workflow recovery, report limits | Plan-mode retry/skip, retained partial files and all-path compile barrier | Complete failed-path delivery/recovery guarantee | Explicit skip-state/barrier semantics and provenance-linked recovery evidence |
| Source lineage and semantic acceptance are not demonstrated | SRC-1; MEM-OBJ-raw, MEM-OBJ-frontmatter, MEM-RTE-refresh-sources, EPI-RTE-thesis-verdict | Source links, conversation exception, refresh overwrite, model-created dates and thesis doctrine | Lossless reconstruction, source truth, checked entailment or an actually accepted novel claim | Versioned source/output histories, checked interpretations and explicit scoped evidence-consuming acceptance decisions |
| External authority and opaque continuation are outside inspected implementation | SRC-1; RT-CMP-host-model, RT-RTE-query-wiki, RT-RTE-adapter-execute, MEM-OBJ-session-digest | Shipped launchers/hooks/helpers, not host scheduling, permission enforcement, compaction internals or private adapters | Host-wide read-only enforcement, transcript preservation, fixed weights, independent remote safety or complete confidentiality | Host grant/parser and continuation contracts, private adapter evidence and independent remote checks; field-specific privacy evidence |
| Persistent updates do not establish learning benefit or exercised self-improvement | SRC-1; MEM-RTE-extract-lessons, RT-RTE-specialist-enable, EPI-RTE-thesis-verdict, RT-RTE-query-evaluation | Durable context and editable guidance paths; no instantiated criticism-to-live-update-to-dependent-operation chain | Improved future capacity from criticism, demonstrated reflection, autonomous theory building or exercised/successful self-improvement | Separate condition-specific evidence, traced organizational adoption and later dependent operation; outcome comparisons for successful improvement |