OpenKB synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Documents become persistent inputs to later work
OpenKB is a document-to-wiki knowledge system with shipped CLI, REST and browser drivers, not one continuously running autonomous agent. This analysis covers its production, maintenance and later-consumption loop at commit ff54396e575ee6feb0113b631a34caa082b441cc, with applicability cutoff 2026-11-09. Its sole source, SRC-1, supplies implementation and doctrine/design. No target execution, actual user KB, model interaction trace or causal experiment supports operational or quality conclusions. External model providers, PageIndex internals and enclosing agent runtimes are outside the frozen implementation boundary. See the boundary and source register.
The central implemented contribution is accumulation through retained text. A user, upload or watched file initiates conversion and compilation. Python schedules model calls and owns filesystem effects; models propose summaries, page content and create/update/related plans. Short-document compilation receives full converted source text. Long-PDF compilation instead uses a PageIndex summary and generated overview for concept/entity production; exported page text remains separately retrievable. A cloud extraction failure or empty response can fall back to local PDF extraction. Neither branch establishes faithful extraction or comprehensive synthesis. RT-RTE-document-add, MEM-OBJ-source-content, EPI-OBJ-summary.
Compilation retains summaries, concepts and entities as Markdown. Selected old concept/entity bodies enter update calls that request full-page replacement integrating new material. Planning receives compact descriptions, fallback body excerpts and entity source counts rather than every existing body. Page generation uses separate explicit contexts. Thus knowledge can accumulate across documents without local weight updates or one ever-growing foreground conversation. This is implemented revision and consolidation, not evidence that the generated text is faithful, novel or true. RT-OBJ-wiki-pages, MEM-OBJ-page-frontmatter, EPI-OBJ-concept, EPI-OBJ-entity.
Admission is principally automatic after workflow initiation. Parsers, type restrictions, nonempty-body checks and link filtering constrain writes, but user initiation is not per-page human approval. The add coordinator requires an ingest lock and attempts snapshot rollback on propagated exceptions. An unparseable plan can nevertheless update the index and return normally, and collected page-generation failures can be omitted while successful pages proceed. Registry/hash admission therefore does not certify semantic completeness. This transaction protocol does not extend automatically to recompile, remove or chat writes. RT-RTE-document-add, MEM-OBJ-hashes.
Retained material reaches different consumers with different force
A later question constructs a query agent with wiki text, page-range and image tools. The model chooses reads and synthesizes an answer; the SDK continues up to a 50-turn cap. Index descriptions and source pointers help navigation, while retained bodies supply evidence through model interpretation. The prompt requests wiki-only, cited answers and narrow source ranges. Those are policies, not semantic or context-volume guarantees: the plain text reader accepts an in-root source JSON path and returns the entire file. This alternate route is reachable in the implementation, not an observed model bypass. RT-RTE-wiki-query, MEM-OBJ-index, EPI-CLM-grounded-answer.
Optional query saves and chat notes create explorations available to later explicit reads. Neither inspected query-save implementation updates the catalog, so retention does not establish automatic rediscovery through the index. The helper and actual CLI save branches also differ in locking, uniqueness and question escaping; the helper's protections cannot be assigned to both. MEM-OBJ-explorations.
Chat has a more direct continuity path. Code selects a session by ID and supplies its sanitized SDK history plus the new message. Sanitization replaces image data URLs with reload hints; it does not semantically summarize the remaining text/tool records. A future model can request the image again. Browser display arrays use a different filter and are not the history supplied to the model. These automatic trace-fed writes and later context consumers are wired; they establish reusable experience-derived context, not demonstrated improved capacity. Saved conversational results provide another retained-text chain, although actual reuse is unobserved. RT-RTE-session-chat, MEM-OBJ-session-history, MEM-OBJ-session-display, MEM-OBJ-explorations.
Instruction loading is distinct from evidence retrieval. Local wiki schema text enters compiler/query/creator instructions. Chat automatically receives discovered skill descriptions and can request a body into the same foreground context; a generic skill runner receives the selected body as instructions. First-hit discovery allows earlier roots to override same-name bundled skills. Generated packages under output/skills/ are not automatically installed in those roots. A nested creator query, by contrast, runs in a fresh context and returns only an answer, not its working history. It does not forward the creator's credential bundle, so top-level credential isolation cannot be extended universally to descendants. RT-BAP-schema-instructions, RT-BAP-skill-instructions, RT-RTE-skill-create.
Maintenance changes availability, not necessarily warrant
Source removal updates source lists and selected links and can delete pages whose last source disappears. It does not regenerate surviving multisource prose to subtract every contribution of the removed document. Explorations, histories and exported skills can independently retain its influence. Workbench/API body replacement likewise preserves existing frontmatter, allowing descriptions or source lists to diverge from edited prose. Recompile is not established as clean-room reconstruction because selected existing bodies can remain inputs. EPI-RTE-source-withdrawal, MEM-RTE-page-edit, RT-RTE-wiki-maintain.
The system does check content, but checks have different consequences. Semantic lint requests contradiction, coverage, staleness and redundancy judgments and retains an advisory report; its inspected drivers do not automatically repair semantic findings. Structural link cleanup can remove a citation while retaining the displayed assertion. Skill evaluation separately compares description-only triggering against generated or supplied labels and asks whether body/reference prefixes support selected questions. It does not execute the skill's task, supply independently verified factual answers or select an installed successor. Reference previews are limited to 4000 text characters per file, and ambiguity/error handling affects score denominators. EPI-RTE-semantic-lint, EPI-RTE-link-cleanup, EPI-RTE-trigger-grade, EPI-RTE-body-grade.
These mechanisms support compilation, diagnostic checking and operational retention. They do not support the stronger assurance that synthesis reflects everything consumed or that cited answers are reliably grounded. Imported source warrant remains unknown. Without candidate instances and their evidence, generated propositions cannot be classified as faithful reshaping, checked entailment or ampliative conjecture. The inspected compiler/query/lint/evaluation paths do not establish evidence-consuming semantic acceptance with a named proposition scope, criterion and intended use; their indexing is consequently not established as integration after such acceptance. This is a bounded route finding, not an absence claim about every helper or opaque model process. EPI-CLM-current-synthesis, EPI-CLM-grounded-answer, EPI-CLM-body-alignment, epistemic conclusion.
Representative delivery: a deck and an optional critic
The selected multi-step workflow is CLI deck generation with critique. Python fixes the order: preflight and authorized replacement, producer run, conditional structural validation, then a fresh critic run. The default producer receives the wiki schema, selected skill body, brief and templated output path. The critic receives the artifact path and original brief, not producer history or validation diagnostics. Shared files carry effects between isolated contexts. The producer has 120 turns with critique enabled, and the critic its own 40; these are independent caps, not one shared budget. RT-RTE-deck-produce, workflow account.
A missing expected producer file raises an error. Structural validation records errors without preventing the inner wrapper's continuation; the outer CLI later exits nonzero on those errors while leaving files on disk. Critique occurs after validation, and the wrapper returns the producer result without revalidating final bytes. Missing critic is tolerated; other failures need not be. Authorized replacement retains a prior copy, but automatic restoration after partial failure is not established. The critic also has conflicting trimming and no-body-edit directions, so content preservation cannot be inferred from its prompt. File retention, command success and final-artifact validity are separate outcomes. WF-CLM-cli-delivery, RT-RTE-deck-produce.
Assessment relative to the work performed
For repeated consultation of an evolving document collection, OpenKB has implemented accumulation, selective read-back, source reload and multiple maintenance paths. For tasks requiring proposition-level retraction, comprehensive source coverage, verified answer truth or validated final post-critic output, those same paths do not provide the required guarantees. The external navigation skill adds an afforded direct-reading route, but its untrusted-data policy does not establish host enforcement or prompt-injection immunity. Shared write-helper path checks constrain that helper under a stable filesystem; they are not an OS sandbox or a restriction on every writer. MEM-RTE-host-read, RT-CLM-wiki-trust, RT-RTE-session-chat.
The strongest learning-related result is durable context reuse and revisable knowledge/instruction artifacts. Improved capacity for future action remains unestablished: no comparison attributes a capacity gain to these updates. The optional critic changes an external work product, not the installed producer method. Its wrapper does not establish content-directed criticism of a consumed method, retained method revision or a next round shaped by that revision. Theory-builder conditions 1–4 are therefore not established as a complete route; the records do not warrant converting inaccessible model reasoning into either formulated criticism or its absence. Reflection is independently unestablished because self-representation with both causal directions is not shown. Model tool selection within bounded requests is computational initiative, not evidence of an autonomous theory builder. Exercised self-improvement is independently unestablished: neither retained context nor exported methods close the evidence-responsive own-organization-change and subsequent causal-use chain. RT-RTE-deck-produce, RT-RTE-skill-create, epistemic conclusion, profile verification.
Concrete changes that would alter these assessments include a semantic acceptance gate with explicit evidence and consequences; claim-level dependency tracking and downstream withdrawal; post-critic validation; or automatic installation and later use of an evidence-responsive method revision. Evidence would need to distinguish those implemented capacities from exercised behavior and benefit: source-to-output traces, inspected final artifacts, resumed-session consumers and objective-specific outcome comparisons. A component benefit claim would require independent variation of that component, not merely a successful whole-system example.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Implementation and doctrine are not observed operation or efficacy. | SRC-1; RT-RTE-document-add, RT-RTE-wiki-query, RT-RTE-deck-produce | Frozen shipped production, consumption and delivery paths; no target execution | Successful use, reliable compliance, learning benefit, quality or causal component effects | Provenance-linked execution traces and outputs; objective-specific comparisons with appropriate component interventions |
| Provider/PageIndex internals, hosted persistence and immutable model identity remain unresolved; external hosts and built distribution bytes are not inspected. | SRC-1; boundary, RT-CMP-pageindex-client, RT-CMP-litellm-model, MEM-RTE-host-read | Local call sites, exported artifacts and shipped host instructions | Complete private-memory classification, extraction fidelity, provider guarantees, deployed-host compliance or behavior of a particular wheel | Pinned dependency/service contracts and implementations, extraction comparisons, host traces and identified distribution artifacts |
| Source warrant and generated proposition relations are unresolved; structural admission is not semantic acceptance. | SRC-1; EPI-OBJ-summary, EPI-OBJ-concept, EPI-OBJ-entity, EPI-RTE-body-grade | Compiler/query and diagnostic evaluator routes | Warranted novel knowledge, checked derivation, reliable factual answers or task capability from grader scores | Candidate-level source support, source-truth assessment, explicit acceptance decisions and executed-task evidence |
| Withdrawal and routing metadata do not track all semantic influence. | SRC-1; EPI-RTE-source-withdrawal, MEM-RTE-page-edit, MEM-OBJ-index, MEM-OBJ-explorations | Source cleanup, body replacement, catalog and two exploration saves | Complete factual forgetting, automatic metadata/body equivalence or guaranteed exploration rediscovery | Claim-level lineage and descendant invalidation; metadata refresh and catalog admission mechanisms with consumer traces |
| Initial local-instruction control and installation/activation are unresolved. | SRC-1; RT-OBJ-agents-md, RT-BAP-skill-instructions, MEM-OBJ-skill-package | Known read/refresh/discovery/export routes | Complete initial authorship/admission account or generated-skill behavioral effect | Provenance and approval records plus installation and later-consumption evidence |
| Deck checks precede optional mutation, and critic instructions do not settle preservation. | SRC-1; RT-RTE-deck-produce, WF-CLM-cli-delivery | Default CLI producer/check/critic path | Final-byte validity, content preservation, critic benefit or automatic recovery | Post-critic validation, artifact diffs, resolved critic contract and recovery traces |
| The partial authority inventory omits the exploration-producing update consumer as a separate learning-authority path. | SRC-1; MEM-OBJ-explorations, RT-RTE-session-chat, profile verification limit | Accepted records and partial comparison inventory; history independently supports learning authority | Complete comparison of which trace-fed update consumers carry learning authority, not existence of that authority or later knowledge consumption | A separately classified exploration-update consumer grounded in these retained routes |
| Partial form and lineage inventories do not separately classify slide prose or lint-report authorship. | SRC-1; RT-RTE-deck-produce, WF-CLM-cli-delivery, EPI-OBJ-lint-report, profile verification limit | Deck markup and diagnostic prose are covered, but not all output parts separately | Complete part-by-part output form/derivation comparisons; independently supported natural-language and authored existence findings remain valid | Separate classifications of retained slide prose and derived lint assessment against the accepted records |