WikiBrain synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
WikiBrain is a persistent knowledge system with built-in compilation, chat and maintenance agents, plus tools for external AI clients. This account covers the shipped repository at commit 32cc135877c8d7cd7be997ce57f25bb83128d891, with an analysis cutoff of 2026-11-08. Its evidence is inspected implementation and declared design from SRC-1, not a running hosted service. The whole-system boundary includes server jobs, browser callers, imports, templates and later consumers; it excludes dependency internals and private external-client orchestration. The optional private operator console is unavailable. No model transcript, user wiki, completed recovery drill or experiment was supplied, and no target code was executed. See the boundary and Source register.
Capture first, compile in a separate job
The system separates raw/ sources, wiki/ compiled pages and schema/ rules. These are logical page paths in PostgreSQL, not three filesystem stores. Markdown prose carries source assertions, knowledge and instructions; paths, versions, parsed links and tags have fixed software consumers. Source location does not certify origin: imports capture converted text and metadata, while direct clients can also create content. Some conversion branches substitute an abstract or retain bibliography instead of full text. The raw layer protects stored conversion output against ordinary normalized updates; it does not establish complete original capture or source truth. Manual import and configured hourly Zotero synchronization are separate admission mechanisms, with the latter requiring no per-item human decision. See MEM-OBJ-raw, MEM-RTE-source-admission and EPI-RTE-source-import.
Compilation starts separately when a user requests ingest. Explicit existing workspace paths are accepted; otherwise code selects pending raw sources. Pending means that no live wiki page links to the source's note ID, subject to archive and special-page exclusions. A backlink can therefore remove a source from the default queue without demonstrating that it was adequately compiled. See RT-RTE-ingest-job, MEM-OBJ-links-tags and EPI-RTE-pending-selection.
The representative workflow is this server-side compilation request, not an invented sequence joining import, chat and lint. The server supplies target paths, optional guidance, provider configuration, tool schemas and a workspace-bound executor. The remote model chooses the next reads and writes among six note tools. General English rules request source reading, related-page inspection, summaries, concept/entity updates, index/log maintenance and explicit conflict marking. These are instructions, not executor-checked completion criteria. There is no separate critic or mandatory human acceptance between generation and ordinary page admission. See the workflow account and RT-RTE-ingest-job.
The built-in provider loops execute chosen tools sequentially and return full results to the next model call. They isolate each job's working messages but accumulate its turns and tool results without local trimming, summarization or delegated-worker contexts. Tool-call caps do not bound total input tokens. Provider/model identifiers are configurable, not immutable weight pins; remote endpoint behavior and parameters remain outside inspection. See RT-CMP-provider-models and the memory context annotations.
Persistence makes results reusable before it makes them warranted
Successful writes persist each note, snapshot, links and tags in an individual transaction. Optimistic version checks can reject stale updates and return current content for re-editing. They do not check factual support or require the requested multi-page products. A job may fail after committing useful or defective partial changes. Runner termination produces done, not verified compilation; the browser refreshes pages after either success or failure. Persisted logs clip results to 400 characters and retain only the last 200 events. Their permissive error parsing can make displayed write summaries unreliable as proof of successful writes, even though the active model receives full results. See RT-RTE-ingest-job, RT-OBJ-ingest-jobs and the workflow delivery account.
Later agents retrieve current notes through exact reads, lists and lexical search. Search ANDs up to eight terms across title/body, can restrict folder/tag, orders newest first and returns short snippets. The prose index provides a separate model-interpreted catalog. Index-first and rule-first behavior are requested, not forced. get_instructions returns all live schema pages in path order, plus pending-source hints, without an aggregate token budget. That response is requested retrieval, not compulsory preloading. Default ingest targets, by contrast, are code-selected metadata supplied to the model task before it requests source bodies. These distinctions matter when assessing what actually reaches a consumer. See MEM-RTE-note-retrieval and RT-BAP-schema-instructions.
Chat supplies another reuse path. New turns automatically receive the selected session's saved user/assistant prose, including failed replies, but not previous tool bodies. Chat has write tools; it is not a read-only query validator. A separate browser filing action copies a selected assistant answer into a wiki query page. Role eligibility and user selection do not establish truth or an evidence-consuming acceptance criterion. See RT-RTE-chat-session and MEM-OBJ-chat-sessions.
The strongest supported experience-reuse contribution is more specific than saving conversations. A browser action can send a selected discussion's nonfailed message text, limited to its last 12,000 characters, into compilation guidance. Models then choose durable wiki writes without per-page approval, and later readers can retrieve those pages. Separately, recorded tool events can produce written-path fallback text retained as an assistant reply and delivered on a later turn. These are wired automatic trace-to-artifact-to-consumer chains, using session logs and tool traces respectively. Ordinary history copying alone does not establish such a derived update. The profile verification accepts these local trace-learning findings while leaving other history and generic caller branches unresolved. Their warrants are MEM-RTE-discussion-guidance, RT-OBJ-ingest-jobs, MEM-OBJ-chat-sessions and RT-RTE-chat-session. None demonstrates improved future capacity.
Maintenance and rule authority have different checks
Structural lint derives facts about queried links, index membership and heading syntax. It licenses those local structural conclusions, not source truth or semantic consistency. Model-backed lint receives a bounded structural prompt and is asked to sample pages for contradictions, staleness and gaps, repair straightforward problems and record suggestions. The same model chooses repairs; there is no independent semantic acceptance or mandatory repair recheck. Without concrete inputs and outputs, compilation and semantic diagnosis cannot be classified as faithful reshaping, checked derivation or novel conjecture merely from page labels. See EPI-RTE-structural-check, EPI-RTE-semantic-check and RT-RTE-lint-job.
Rules themselves are mutable retained guidance. Template updates compare delivered bases, current rules and shipped text; safe mode retains identified conflicts, while overwrite can replace them. Browser overwrite confirmation is not a backend human-only requirement: authorized REST callers can select the mode. Comparison bases and editable custom bundles are not active instructions until applied to live schema. Moreover, comparison and eventual version rereading are separated, so this policy does not guarantee preservation of every intervening concurrent edit. See RT-RTE-rule-update, MEM-OBJ-template-state, MEM-OBJ-user-templates and the report verification limits.
The declared human-owned rule boundary is also narrower in implementation than its wording suggests. Built-in create/update dispatch denies literal schema/ prefixes before note-path normalization. A string such as /schema/example.md can pass that guard and normalize to a schema write, provided remaining quota, content and database checks pass; update additionally needs a valid version. Later instruction retrieval can select the resulting note. This is supported static conditional reachability, not an executed exploit or evidence of a model choosing that spelling. MCP deliberately permits broader write-authorized schema edits and lacks the built-in raw-read untrusted label. Neither a label nor a natural-language instruction guarantees model compliance. See RT-CLM-schema-guard, RT-RTE-schema-spelling and RT-RTE-client-tools.
Assessment depends on the requested outcome
For a scenario requiring retained compilation, later retrieval and localized editing/recovery, the source establishes connected implementation rather than a storage-only API. For a scenario requiring accepted factual claims, exhaustive contradiction detection or verified multi-page delivery, the same paths do not supply that guarantee. Persistence supports the reuse part of the compounding claim; it does not demonstrate compounding epistemic benefit. Structural lint and requested semantic checking support different parts of the health-check claim. See EPI-CLM-compiling-compounds and EPI-CLM-health-check.
The property conclusions remain separate. Durable trace-fed context is wired, but criticism of a consumed theory improving future capacity is not demonstrated. For topic theories, localized theory instances and their actual content-dependent consumption are not supplied; a request for criticism does not establish a challenged claim, blame placement and changed reliance; nor does requested report retention establish criticism shaping a next round. Thus theory-builder conditions 1–4 are not jointly established by these routes, with distinct gaps for each condition. Reflective self-representation is independently unestablished. Automatic internal tool decisions do not establish an autonomous theory builder performing every required internal operation. Mutable rules and later delivery do not close an exercised self-improvement pathway: objective-bearing evidence determining an update and subsequent operation causally depending on that change remain unestablished. These are bounded evidence conclusions, not negatives about opaque model processing or all scenario instructions. See EPI-RTE-semantic-check, the epistemic conclusion and RT-RTE-schema-spelling.
Concrete evidence could change this assessment: linked source/output/transcript instances could establish the semantic transformation and exercised checking; retained criticisms and later decisions could establish iteration; matched outcome comparisons could test benefit at the intervention's actual scope. System changes such as normalized-path authorization before dispatch, a mandatory evidence-consuming acceptance stage, or concurrency-bound template comparison/admission would change the corresponding architectural findings, but would still require separate evidence of operation and benefit.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied execution or causal comparison | SRC-1; RT-RTE-ingest-job, EPI-RTE-semantic-check, EPI-CLM-compiling-compounds | Frozen implementation/design, not hosted operation; no target execution | Observed completion, reliable criticism, improved future capacity, exercised self-improvement and causal compounding benefit | Provenance-linked model/tool/page traces, subsequent consumer behavior and appropriately scoped comparisons |
| Scenario-specific epistemic doctrine is not separately inventoried | SRC-1; MEM-OBJ-schema, RT-RTE-rule-update, RT-BAP-schema-instructions; report verification limit | General English rules dominate the reports; verifier additionally inspected researcher and PM doctrine | Complete scenario-level epistemic-guidance inventory or a system-wide negative about such guidance. Researcher rules request falsifiable claim/evidence/counter-evidence/judgment pages; PM rules request rationale and reversed-decision retention, without observed fulfillment | Accepted records separately inventorying those instructions and their consumers; concrete fulfillment evidence for operation claims |
| Template comparison and admission do not bind the same read version | SRC-1; RT-RTE-rule-update; report verification limit | Snapshot-based state/text computation followed by a later note-version read | Concurrency-safe merge/admission or preservation of every intervening edit; no executed lost update asserted | Comparison-bound version admission or transactional implementation, plus scoped concurrent-operation evidence |
| Capture fidelity and upstream provenance remain bounded | SRC-1; EPI-RTE-source-import, MEM-RTE-source-admission, RT-RTE-client-tools | Selected conversion branches and local admission; not every parser, remote response or caller's content production | Complete-original capture, source truth, exhaustive lineage or universal human content approval | Original/captured pairs, parser-specific inspection and provenance/admission records from actual callers |
| Consumer projections omit evidence and reasons | SRC-1; MEM-OBJ-chat-sessions, MEM-RTE-discussion-guidance, RT-OBJ-ingest-jobs | Prose-only later history, clipped job events and transcript tail; no compulsory later reading of log reasons | Complete episode reconstruction, reliable write certification from logs or guaranteed rationale survival | Full provenance-linked episode artifacts and a later consumer trace showing which reasons and results were received |
| Rule protection is conditional, and activation is unobserved | SRC-1; RT-RTE-schema-spelling, RT-CLM-schema-guard | Static original-string guard and normalized note functions; remaining admission checks still apply | Universal automatic schema-write exclusion, an observed exploit or demonstrated downstream policy change | Canonical authorization covering accepted spellings; for operation claims, a trace through admitted write and later behavior |
| Withdrawal and recovery are consumer-relative | SRC-1; MEM-RTE-state-withdrawal, MEM-OBJ-note-versions, MEM-RTE-human-delivery | Per-page rollback; 7/90-day non-current snapshot windows; archive, soft deletion/reset and separate chat/export copies | Job-wide undo, physical erasure, automatic retraction of dependent claims or guaranteed restore outcome | Dependency/copy-aware withdrawal implementation, end-to-end deletion inspection and completed restore evidence |
| Visual payload and external behavior remain unresolved | SRC-1; MEM-OBJ-assets, MEM-RTE-human-delivery, RT-CMP-provider-models; boundary | Binary browser/share/export paths versus Markdown-only built-in tools; provider/client internals and private console outside available evidence | Built-in binary model perception, exhaustive visual-form classification, immutable model identity or private deployment/compliance claims | Inspected model-image input wiring, pinned model artifacts and authorized evidence for the excluded consumers/deployment |
| Trace-update inventory is partial | SRC-1; MEM-RTE-discussion-guidance, MEM-OBJ-chat-sessions, RT-RTE-client-tools; profile verification | Specific discussion and fallback positives; ordinary-history derivation and opaque caller/input alternatives unresolved | Exhaustive trace-learning/source inventory or system-wide absence; local positives do not prove benefit | Records isolating each remaining automatic derivation, its original inputs, durable output and later consumer |