VLM-wiki synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
VLM-wiki is a personal knowledge workflow, not a general agent host. It asks an external AI host to compile media into a persistent Markdown wiki and answer later questions from it. Alongside that procedure, it ships independent Python tools for import, image analysis, historical review, work-project reporting and notification. The distinction matters: implemented report writers do not establish an implemented end-to-end model compiler. The evidence is SRC-1, https://github.com/veniVeci/VLM-wiki at commit 503c43f98b5d1e21c74df384fdffd3e0652573b3, with analysis cutoff 2026-10-14. The whole-system boundary includes shipped callers, instructions, maintenance and delivery; host/provider internals, private deployment state and external services are outside it. This is code-grounded source analysis, with doctrine and attributed examples but no supplied execution trace or causal experiment. See the boundary and evidence and Source register.
From inputs to retained descriptions
A user can place media for the host to interpret or invoke the importer to copy configured local sources. Import is implemented separately from compilation. Python admits files using source configuration, destination rules and a persisted path/size/integer-modification-time fingerprint. Later imports reload that manifest and skip matching keys. This is a concrete cross-invocation control loop, but not content deduplication: another path can produce another import, unchanged metadata can conceal changed bytes, and deleting a retained copy does not clear its skip key. Copies precede the eventual manifest save, so interruption need not leave consistent state. See RT-RTE-source-import and MEM-OBJ-import-manifest.
The direct image drivers send one image and a fixed descriptive prompt to Qwen, then persist returned prose without per-response factual review. They do not send the wiki or carry earlier image responses into the next request. Their model and endpoint choices bypass the general host-model configuration; neither the Qwen alias nor the declared host selection pins immutable model weights. Batch confirmation authorizes starting calls, not accepting their claims. The batch catches inference/report-write errors per image and continues, but its final summary write is outside that handler. A completed batch therefore need not mean every image succeeded or every artifact survived. See RT-RTE-image-analysis, RT-CMP-qwen and the workflow failure account.
Supporting tools sample video frames at uniform indices and measure image dimensions, format, mode and file size. They establish preparation and physical measurement, not scene-change detection, transcription or caption verification. Import places images in month folders while the batch image selector reads only its immediate directory; these tools cannot be silently joined into an automatic compiler. See MEM-RTE-frame-extraction, MEM-RTE-metadata-extraction and RT-RTE-image-analysis.
From media interpretation to later answers
The representative model-dependent workflow is instructed image ingest followed by a separately requested query. The user supplies media; the host supplies model execution and file tools. Instructions ask for visual interpretation, metadata extraction, captions, category-page decisions, related updates and persistent articles. These stages are not established separate agents or isolated contexts. Later, a question should cause the model to read the index, select articles, prefer their content over training knowledge and answer with internal links. Saving that answer as an archive requires another user request. These are supported procedures, not observed compliant host behavior. See RT-RTE-agent-workflow, RT-RTE-wiki-query and the workflow account.
The semantic memory is local prose with source pointers, links and metadata. Source pointers afford inspection; they do not preserve every derivation or establish acceptance. Index-led loading selects external files rather than implementing conversation compaction, and no token budget or host context policy is established. Navigation also has concrete weaknesses: some tracked index targets lack pages, and generated project links can disagree with slugified filenames. Thus persistence alone establishes neither discoverability nor arrival in a later answer. See MEM-OBJ-wiki-articles, MEM-OBJ-wiki-index and MEM-CLM-index-targets.
From retained history to reports and delivery
The historical-review engine implements reuse without a model. It collects direct Markdown files across seven wiki categories and direct diary Markdown, then derives counts, date order, selected clues and fixed recommendations. Daily and weekly outputs use the same collected corpus, not distinct date windows. Its limited frontmatter parser can lose media structure, month-only dates fail latest-date parsing, and article bodies are not semantically summarized. Numerical derivation is supported over the parsed representation, not over a complete or independently verified life history. See RT-RTE-history-review, MEM-CLM-parser-projection and EPI-RTE-history-statistics.
The work cycle instead scans configured external project files and Git metadata, writes project pages, replaces the Projects index section, writes a review and optionally starts WeCom delivery. Project pages are eligible for a later historical review even if they subsequently lose their index entries. The sender rereads the review and selects at most twenty cleaned preview lines, capped at 4,000 characters. This is producer-initiated push to human recipients, not unsolicited model context. The documented standalone historical-review send has a different selection outcome: its English preview heading fails the sender's Chinese-heading test, causing fallback to the first twenty cleaned lines. That mismatch does not prove all recommendations are omitted or the service call fails. See RT-RTE-work-cycle, MEM-OBJ-project-pages, MEM-OBJ-work-review and MEM-CLM-daily-delivery.
Delivery and durable completion are separate. Subprocess failure stops the cycle, but earlier page/index writes are not rolled back. Empty activity can leave only one recommendation where rendering indexes three, after index replacement. Standalone producer/sender and skip-push routes further separate local output from transmission. Windows wrappers supply launch/logging entry points, not installed recurrence or reliable success evidence. These are source-level failure possibilities, not observed incidents. See RT-RTE-work-cycle.
What the reuse establishes
The strongest supported learning-related contribution is an implemented activity-to-durable-page/advice pathway with later history-review or sender consumers. It qualifies locally as automatic trace-fed memory writing: external activity and Git history shape retained material capable of affecting later reports or human decisions. The accepted profile verification preserves that positive separately from unresolved article/diary origins and missing readers. Unchanged Python rules do not negate this contribution. Conversely, that pathway does not demonstrate that criticism improved the system's capacity for future action. Latest JSON briefs, analysis prose and logs have no established later semantic consumer merely because writers persist them. See MEM-OBJ-project-pages, MEM-OBJ-work-review and MEM-OBJ-runtime-briefs.
For personal recollection, the supported contribution is instructed source-linked accumulation and query. For activity awareness, the stronger implementation evidence is deterministic reporting and optional preview delivery. Neither scenario establishes warranted explanatory conclusions. A positive place counter triggers a single-trip-bias assertion even for multiple trips; any nonempty work activity can trigger a fixed named-domain attention claim regardless of content. These are implemented conjectures alongside narrower measurements, not accepted truths. Retention, configuration gates and HTTP success do not supply an evidence-consuming acceptance decision. See EPI-RTE-history-interpretation, EPI-RTE-work-interpretation and EPI-CLM-history-synthesis.
Maintenance changes reliance labels but does not close that gap. The disclaimer helper checks a marker and rewrites matched reports in a particular directory as simulated; it does not verify their provenance or propagate withdrawal to dependent articles. Raw immutability is a host/user convention, not filesystem enforcement over the raw tree. See EPI-RTE-disclaimer-rewrite and RT-CLM-raw-immutability.
The accepted runtime does not establish a formulated-theory route with separate findings for theory-builder conditions 1–4, so this account assigns none from writing, persistence or fixed-rule execution alone. Content-directed criticism improving future capacity remains unestablished. Reflection is independently unestablished: historical subject matter and the name “reflection” do not establish bidirectional self-representation. Fixed computations own their next steps once invoked, but autonomy of the host-mediated workflow or an autonomous theory builder is not established. Self-improvement is likewise unestablished at this repository boundary: activity-fed external reports do not by themselves show evidence-responsive change to the system's own organization followed by operation causally dependent on that change. These are separate limits, not a bundled negative or a judgment about opaque model internals. See RT-RTE-agent-workflow, RT-RTE-history-review, RT-RTE-work-cycle and the epistemic conclusion.
Host traces linking media, admitted pages, retrieval and answers would change the execution assessment. Reference-grounded checks and recorded claim dispositions would change the acceptance assessment. Repaired parser/navigation/preview rules would change fidelity at those specific boundaries. Actual scheduler and delivery traces would change recurrence/receipt findings. An independently specified improvement objective, evidence-responsive internal update, consequential later consumption and comparative outcomes would be needed to assess exercised self-improvement and benefit. Neither verification declared additional limits or blockers; see report verification and profile verification. Source-read log: none; this synthesis uses supplied members only.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No execution or causal comparison | SRC-1; RT-RTE-agent-workflow, RT-RTE-work-cycle | Shipped source and attributed artifacts at the frozen commit | Successful compiler operation, longitudinal reuse, delivery, user benefit or causal improvement | Inspectable end-to-end traces and appropriately scoped outcome comparisons |
| External host and provider behavior | SRC-1; RT-CMP-ide-model, RT-CMP-qwen, RT-RTE-wiki-query | Client requests and host-facing instructions, not host/provider internals | Actual instruction loading, grants, fallback, context management, model fixity or parameter mutation | Host configuration/traces and provider version/mutation evidence |
| Private inputs and generated state unavailable | SRC-1; RT-OBJ-config, MEM-OBJ-runtime-briefs, RT-RTE-work-cycle | Tracked configuration, writers/readers and wrappers; ignored outputs and external trees unavailable | Deployed capability envelope, actual recurrence, real input coverage or reconstruction of a deployed cycle | Frozen local overrides, manifests, outputs, external input snapshots and scheduler/service records |
| Unknown instance provenance and unaudited media meaning | SRC-1; MEM-OBJ-analysis-reports, MEM-OBJ-wiki-articles, MEM-OBJ-raw-inputs | Retained text and declared producer paths; pixels not audited | Authenticated sample authorship/admission, caption accuracy or complete trace origins | Linked input/output and admission traces, reference evidence and media-content audit |
| Persistence is not complete later consumption | SRC-1; MEM-OBJ-runtime-briefs, MEM-OBJ-analysis-reports, MEM-OBJ-wiki-log | Shipped object inventory and established readers | Automatic report-to-wiki compilation, brief feedback, or complete downstream authority/form classification | A source-supported later reader and evidence of the payload it receives |
| Lossy projection and inconsistent navigation/delivery | SRC-1; MEM-CLM-parser-projection, MEM-CLM-index-targets, MEM-CLM-daily-delivery | Current parser, tracked index/pages and sender selector | Faithful full-corpus synthesis, complete article discovery or selection of the designated historical preview | Corrected rules plus parser, link and delivered-payload checks |
| Candidate generation and provenance labels lack semantic acceptance | SRC-1; RT-RTE-image-analysis, EPI-RTE-history-interpretation, EPI-RTE-work-interpretation, EPI-RTE-disclaimer-rewrite | Direct persistence, fixed inference branches and marker rewrite | Caption truth, validated explanations, authenticated simulation attribution or dependency-wide withdrawal | Evidence-consuming checks against named criteria, recorded dispositions and traced dependent updates |
| Durable updates do not establish improved capacity or self-improvement | SRC-1; RT-RTE-history-review, RT-RTE-work-cycle, MEM-OBJ-project-pages | Fixed review/scanner code and later-consumer wiring, excluding opaque model processing | Criticism-driven capacity gain, reflection, autonomous theory-building or exercised self-improvement | Separately supported property records, retained criticism/update lineage, consequential later operation and objective-relative outcomes |