Quicky Wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Documents become reusable graph context
Quicky Wiki is a document-to-knowledge compiler with later question answering, not a general model-controlled agent loop. Host code schedules extraction, comparison, graph mutation, page assignment, summaries and answers. Models supply semantic judgments and prose; host code owns filesystem, database and network effects. The shipped CLI, library, browser dashboard, watcher and MCP interfaces expose different entrances to that machinery. An enclosing agent's planning is outside this account (RT-RTE-source-ingest, RT-RTE-client-access).
The evidence is source implementation and shipped documentation/prompts from https://github.com/anzal1/quicky-wiki at cfeecbffce5b6d0f16533b258a1e88edd86f731e, with applicability cutoff 2026-10-15. The whole-system boundary includes production, retention, maintenance and consumption, not just the query API. No target execution, corpus, generated database, provider response, operational trace or causal experiment was supplied. Thus implemented paths below are not observed successful operations. SRC-1 and its exclusions are defined in the boundary and Source register; both report verification and profile verification found no remaining blockers or additional declared limits.
A representative shipped workflow starts with a local-file qw ingest. Setup and later command loaders select configuration and reopen .quicky/graph.sqlite. Ingestion hashes file bytes, separates frontmatter and body, resolves entity kind/metadata, and registers the source or new hash before model work completes. Extraction requests atomic factual statements, confidence, tags and related concepts. Existing statements and new candidates then reach a comparison model, unless the empty-store shortcut makes all candidates new. Parsed actions automatically reinforce, challenge or add claims; initiation is not per-claim human approval. Page assignment and summary generation follow, then the CLI renders Markdown. Compiler return, early success text and completed file delivery are separate milestones (RT-RTE-runtime-setup, RT-RTE-source-ingest, RT-OBJ-knowledge-diff, workflow account; SRC-1, src/compiler/ingest.ts, src/compiler/resolve.ts, src/cli/ingest.ts).
This progression is availability-oriented rather than a sequence of independent factual approval gates. Malformed extraction can produce no claims; malformed comparison can admit every extracted claim as new. Omitted matches can omit candidates. Assignment has routing defaults, and summary failures are noncritical. There is no whole-ingest transaction or retry coordinator. A failed attempt can leave a registered hash, earlier graph writes or some rendered files; rerunning unchanged bytes can skip unfinished extraction/resolution. Method-local SQLite transactions do not cover the complete workflow, and empty-page prune itself lacks a method-wide transaction (RT-RTE-source-ingest, RT-RTE-health-review, workflow account; SRC-1, src/compiler/diff.ts, src/graph/store.ts).
Retention changes what later calls receive
The durable memory separates natural-language claims and summaries from symbolic scores, identities, provenance, relations and mutation events. Source rows retain paths, hashes and labels, not immutable source-body snapshots. The authoritative query substrate is SQLite; rendered Markdown is a separate derived file surface. Statements and scores reach later models as evidence-like context, while edited entity guidance enters extraction's system message as instructions. These are different authority paths, not uniformly enforced beliefs (RT-OBJ-knowledge-store, MEM-OBJ-source-register, RT-BAP-retained-context; SRC-1, src/graph/store.ts, src/compiler/ingest.ts).
Working context is reconstructed per call, not maintained as one conversation with global compaction. Extraction above 12,000 characters uses sequential 8,000-character chunks with 500 overlap, then first-wins filtering on normalized 80-character statement prefixes. That filter can discard different propositions sharing a prefix; it is not checked semantic equivalence. Comparison sees all existing statements, assignment all page titles, and summary generation one page's claims. These selectors have no shared aggregate token budget and do not restore hidden reasoning (RT-RTE-source-ingest, EPI-RTE-chunk-selection; SRC-1, src/compiler/ingest.ts, src/compiler/resolve.ts).
A later ordinary question triggers lexical graph search, limited to 50 results per page/claim category, with up to ten claims expanded from a page-only hit. The answer model receives selected statements and rounded confidence beneath page headings—not source bodies, source identities, event reasons or even claim IDs. Returned IDs are unvalidated, and MCP reduces them to a count. The confidence-and-citation feature is implemented as requested structured output, not verified citation integrity or calibration. Consensus instead supplies all page-associated claims to responders; one successful response or a failed comparison can still yield claims labeled agreed. That label therefore does not guarantee multi-model convergence (RT-RTE-question-answer, RT-CLM-confidence-citations, RT-RTE-consensus-answer; SRC-1, src/graph/query.ts, src/compiler/consensus.ts).
Contradiction, maintenance and reuse have different consequences
A challenge normally preserves the old statement, lowers its score and records a reason. It does not necessarily retain the challenger statement; the next ordinary answer does not receive the event reason. Source-quality multipliers, reinforcement formulas and time decay affect reliance without authenticating evidence or calibrating probability. Reinforcement is a weighted blend and need not raise confidence. Dependency cascade is a conditional capability: ordinary ingestion drops extracted dependency strings rather than inserting claim edges, though embedding callers can populate edges through store APIs (MEM-OBJ-claims, MEM-OBJ-claim-controls, MEM-OBJ-epistemic-events, MEM-OBJ-graph-relations; SRC-1, src/compiler/diff.ts, src/compiler/confidence.ts, src/graph/store.ts).
Red-team criticism, resurfacing questions and discovery diagnoses are advisory returned/displayed proposals, not a wired accepted-revision loop. Structural lint checks graph predicates; prune removes empty pages, not false claims. Numeric decay does persist. Ordinary query has no confidence floor, although default slide selection uses one and explicit topic selection bypasses it. Scores therefore have real selection effects without acting as a universal factual veto (RT-RTE-health-review, EPI-RTE-lint-prune, EPI-RTE-slide-output; SRC-1, src/cli/metabolism.ts, src/cli/lint.ts, src/render/marp.ts).
Withdrawal is similarly uneven. Source replacement does not retract every earlier claim, watcher deletion is ignored, and live-graph deletion does not sweep saved answers or rendered files. Summaries can remain stale after direct edits, deletion or decay. Dashboard page display can reload old Markdown alongside current graph data, even though standard model answering reads SQLite. The advertised living-knowledge mechanism is consequently an incremental maintained graph, not a demonstrated coherent evidence archive (MEM-OBJ-page-summaries, MEM-OBJ-rendered-views, RT-RTE-client-access, EPI-CLM-living-knowledge; SRC-1, src/cli/watch.ts, src/cli/serve.ts, src/render/markdown.ts).
What the change loops establish
The strongest supported contribution to future adaptation is a concrete retained-output path: an operator can save a standard question/answer as a raw source; later explicit ingestion or an active watcher automatically re-extracts it into durable graph content available to later questions. This supports a wired trace-fed update, not improved future capacity. Manual retention and automatic claim admission are separate decisions. Recycled answers supply no independent corroboration, and no exercised loop or comparison establishes benefit (RT-RTE-answer-save, RT-RTE-source-ingest; SRC-1, src/cli/save-answer.ts, src/cli/watch.ts).
The independent properties should not be collapsed into a grade:
- Theory-builder condition 1, localized content: retained factual statements provide inspectable whole units. This does not establish a broader explanatory theory or finer-grained assumption structure (MEM-OBJ-claims).
- Condition 2, consumption: later content delivery and model requests are wired; actual decisions depending on what a formulated theory says are not established by delivery alone (RT-BAP-retained-context).
- Condition 3, content-directed criticism: semantic comparison can propose challenges that change reliance, and event notes can retain reasons. Reliable formulated criticism, including placement of blame and accepted factual revision, is not established by the source-only account (RT-RTE-source-ingest, MEM-OBJ-epistemic-events).
- Condition 4, iteration: changed scores can reach later calls, but retained criticism shaping a subsequent criticism round is not established; ordinary answers omit those reasons and red-team proposals stop at display (RT-RTE-health-review, RT-RTE-question-answer).
- Learning from criticism: no evidence shows criticism of consumed content improving future capacity. Reflection: health statistics represent selected graph properties, but the accepted account does not establish a representation-mediated, criticism-driven revision of the system's own method. Autonomy: code sequences ingestion without per-item approval; this is not evidence that computation performs every internal operation of autonomous theory building. These conclusions remain separate from the wired output-reuse path (epistemic bounded conclusion, RT-RTE-health-review).
- Self-improvement at this boundary across later invocations: graph/configuration changes can affect later operation, but an objective-bearing, evidence-responsive self-change pathway is not established. Save/reingest lacks demonstrated objective-based update determination and subsequent causally dependent operation; configurable prompts are not an automatic critique-to-method update. Neither exercised nor successful self-improvement follows. Adapter calls request inference, not inspected weight training; provider internals remain opaque (RT-RTE-answer-save, RT-RTE-runtime-setup, RT-CMP-llm-adapter).
For document accumulation followed by graph-based answering, the source establishes an end-to-end implemented path. For scenarios requiring factual acceptance, reliable retries, propagated withdrawal, inspected criticism reuse or calibrated answers, it does not establish those stronger properties. Concrete source/output pairs and acceptance traces would change the semantic assessment; dependency-edge producers and coherent invalidation/recovery mechanisms would change the architectural assessment; objective-linked updates with traced later consumption and controlled outcome comparisons would change the learning and self-improvement assessment. These are different evidence requirements, not product rankings.
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-source-ingest, RT-RTE-answer-save, RT-RTE-health-review | Frozen shipped code and prompts, not deployment artifacts | Successful activation, operational recovery, improved future capacity or causal benefit | Linked inputs, provider outputs, database/file transitions and later-consumer traces; objective-specific comparison with appropriate intervention |
| Concrete generated propositions and independent factual judgments unavailable | SRC-1; EPI-OBJ-source-content, EPI-OBJ-query-results, MEM-OBJ-page-summaries, EPI-RTE-slide-output | Generation requests, technical checks and persistence paths | Extraction/summary fidelity, entailed derivation versus conjecture, accepted ampliative claims, calibrated confidence or accurate answers | Source/output pairs, independently warranted references and evidence-consuming acceptance decisions with criterion, use and scope |
| External models, dependencies and deployment envelope uninspected | SRC-1; RT-CMP-llm-adapter, RT-RTE-client-access | Adapter and host dispatch implementation only | Complete retained-state classification, provider learning/fixity, deployed isolation, interoperability or security guarantees | Pinned provider/dependency evidence and deployment grants, network controls and relevant operation traces |
| Embedding initialization and generic caller control unresolved | SRC-1; MEM-OBJ-graph-relations, MEM-OBJ-claims, RT-RTE-runtime-setup | Public APIs and shipped producers; no initialized external database/configuration | Complete provenance/admission inventory or exercised cascade from externally supplied edges | Caller implementations, initialization/admission records and populated-edge execution traces |
| Views and source lifetimes do not share one invalidation path | SRC-1; MEM-OBJ-rendered-views, MEM-OBJ-page-summaries, RT-RTE-answer-save | Inspected graph edits, watcher, summary refresh, rendering and dashboard file reads | Complete withdrawal or identification of what an actual deployment displayed | Versioned artifacts and consumer traces; coherent invalidation/cleanup and recovery implementation |
| Remaining export/UI details and dependency internals are not exhaustively assessed | SRC-1; MEM-OBJ-rendered-views, RT-RTE-client-access, boundary coverage | Positive traced routes and bounded searches, not a complete semantic/security audit | System-complete negative about all export transformations, complete payload classification or exhaustive safety claim | Focused inspection of remaining relevant templates, UI behavior and dependencies, with route-specific records |
| Retained criticism is not established as consumed criticism or an improvement-directed method update | SRC-1; MEM-OBJ-epistemic-events, RT-RTE-health-review, RT-BAP-retained-context | Challenge-event storage, ordinary context construction and advisory review callers | Completed theory-building iteration, learning from criticism, reflective method revision or exercised self-improvement | Named criticism and revision decisions, retained rationale reaching the next round, objective-bearing update determination and causally dependent subsequent operation |
No unresolved inter-report conflict was supplied, and neither verification declares an additional limit. No frozen source reread was needed for this synthesis.