llm-wiki-newsroom synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From external documents to shared working files
llm-wiki-newsroom is an agent-operated knowledge system, not an enclosing agent runtime. It supplies newsroom role instructions, layered wiki contracts, deterministic Python tools and downstream reading interfaces. An external coding-agent host supplies model execution, tool access, scheduling and context construction. This analysis covers the shipped production, maintenance, improvement and consumption paths of https://github.com/alfadur7/llm-wiki-newsroom at commit 8e209cab483da544852d7213f0e25291d06d8a88, with an applicability cutoff of 2026-11-08. SRC-1 is the sole registered source. The evidence is inspected implementation, design instructions, attributed operational reports and bounded samples of supplied experiment artifacts. No target code or services were executed. See the boundary and evidence and Source register.
The principal operating division is between interpreted editorial work and fixed-rule file transformations. The Editor-in-Chief routes requests and coordinates shared work. Reporter writes source reflections and eligible stubs; Columnist writes higher-level narratives; Copy Editor uses deterministic lint; Desk judges qualitative defects. Authoring and review models use unpinned family aliases, while the orchestrator inherits the session model. Their provider internals are outside the evidence boundary. These instructions specify how work should run; the Python adapters do not implement a scheduler enforcing the entire editorial sequence. See RT-RTE-role-delegation, RT-CMP-opus, RT-CMP-fable, RT-CMP-sonnet and RT-CMP-session-model.
Folder ingestion with at least three genuinely new sources is the representative multi-step workflow. A deterministic prefilter matches URL before normalized raw path. Main reads shared index/overview context, then assigns one new source to each Reporter. Reporters read raw material, write source pages and return proposals, links and contradictions. Main deduplicates proposals, re-evaluates cross-source hub thresholds and serializes shared hub/index/log changes. Fresh role contexts keep internal author traffic out of main's window, but persistent file edits cross that boundary. Source-slug ownership is a coordination rule, not an inspected filesystem lock. A Reporter report is necessary for integration even when its file writes succeeded. See RT-RTE-source-ingest and the workflow account.
Retained source pages combine attributed natural-language claims and quotations with symbolic grade tokens, dates, URLs and links. Hubs and higher-level narratives select and interpret that material. Generated graph edges, source maps, backlinks, dependencies and catalogs provide navigation and routing, not a lossless claim archive or independent factual endorsement. For example, graph generation collapses repeated target links and can promote relation and grade separately. See MEM-OBJ-source, MEM-OBJ-hub, MEM-OBJ-narratives, MEM-OBJ-graph and MEM-OBJ-access-indices.
Checking, revision and publication are separate steps
Source self-lint and rebuild precede applicable qualitative review. Changed higher-layer targets receive their prescribed Desk scope; source review is selected by a threshold of at least seven fact claims and three citations. For originals above 40KB, the Desk contract narrows raw reading to quotation spans and speaker sentences. Desk reports defects rather than editing directly. Authors respond by revision or stated divergence; divergence receives one restricted re-adjudication. A third same-cause failure stops iteration that turn and escalates to the operator. Contradiction consolidation follows relevant claim-text edits and rebuild because claim IDs depend on wording. Final all-group lint closes temporary contradiction-diagnostic exemptions, followed by logs and the completion report. See RT-RTE-source-ingest, EPI-RTE-desk-check and workflow closing order.
The discriminating substantive safeguard is evidence-consuming editorial comparison, not the presence of a grade or a successful lint exit. Citation metrics check syntax and schema. For synthesis, Desk must open every declared source and compare surfaced cross-source joins against component spans. The deterministic join detector only recognizes qualifying single-line declared-source combinations. Targeted lint prints all detected entries as 70-character prefixes plus source slugs; corpus lint computes entries without printing them. Even targeted output is therefore a locator requiring full-page and span reads, not the claim evidence itself. Advisory synthesis checks can return zero despite structural failures. See EPI-RTE-citation-check, EPI-RTE-join-surface and EPI-RTE-desk-check.
These checks support a specified editorial process for wiki publication, not a universal factual guarantee. Checking, deciding publication eligibility and retaining files have different owners and consequences. Pages already exist on the shared filesystem before final review. No production instance was traced from a candidate through an evidence-consuming acceptance decision and then accepted integration. Narrative generation also remains semantically indeterminate: without a claim/premise comparison, neither entailment nor a newly conjectured claim follows from the name “synthesis.” See EPI-RTE-editorial-disposition, EPI-RTE-narrative-authoring and EPI-CLM-newsroom-trust.
The intended completion report, persisted working files and external publication are separate milestones. Commit/push of ingest outputs requires explicit operator approval; automatic queue-file commits are a narrow exception. Direct CLI/export paths and legacy host procedures do not reproduce every newsroom gate. Write/Edit hooks protect selected inputs under the host contract, but shell/CLI effects do not enter that same hook path. Named-tool bans are not capability isolation. These are source-supported scope distinctions, not observed deployment bypasses. See workflow delivery, RT-RTE-build-lint, RT-BAP-role-guidance and RT-BAP-hook-feedback.
Later use selects evidence rather than renewing its warrant
For a later question, the agent procedure classifies the request, retrieves graph/search leads and requests full bodies from at most ten leading pages before answering. QMD's retrieval internals remain opaque; the no-QMD fallback uses graph neighbors and direct reads. An ordinary answer does not write the wiki. Explicit save approval starts a separate reviewed synthesis cycle. Authoring uses a grounding ladder that expands on insufficiency signals, with provisional caps of three consecutive empty results per rung and fifteen expansion operations outside mandatory grounding. At the cap, the author must use whole-corpus reading if it fits or narrow and disclose the evidential shortfall. See RT-RTE-wiki-query, RT-CMP-qmd and MEM-RTE-grounding-ladder.
Other consumers receive materially different selections. The browser reloads copied page bodies; hosted export omits source/entity/concept bodies and most comments, supplying selected narratives and source locators instead. Its estimated token budget recommends uploads rather than enforcing hosted context membership. Email rendering can turn unsupported links into plain aliases, and plaintext drops link targets. Neither remote uploads nor sent mail are automatically recalled when local evidence changes. See MEM-RTE-browser-read, RT-RTE-wiki-export and MEM-RTE-briefing-read.
Maintenance likewise separates access refresh from substantive correction. Rebuild regenerates derived files; it does not itself reground prose. Dependency dates nominate stale narratives without excluding them from ordinary reads. Contradiction derivation carries forward authored assertions and adds heuristic categories and strength, not adjudicated truth. A superseded/open record prompts reconsideration rather than automatic closure. Hub demotion nominates candidates for Desk judgment and link repair, not automatic deletion or universal forgetting. See MEM-OBJ-dependencies, EPI-RTE-contradiction-derive, EPI-RTE-reground-surface and MEM-RTE-hub-withdrawal.
One rebuild path does consume retained history directly: clustering can reuse a prior partition when graph fingerprint and hub coverage match, or initialize native computation from prior memberships. Backend availability and --cold alter those branches; cold does not erase separately loaded label history. This is implemented retained-state dependence, not evidence that historical initialization improves communities. See MEM-RTE-cluster-rebuild.
Experience can shape future work without demonstrated improvement
The strongest supported trace-fed contribution is a durable feedback-review checkpoint. Software extracts conversational reactions, retains derived counts/history and the review window on explicit checkpoint invocation, and consumes that state during later mining. This is a wired automatic trace-fed behavior-shaping routing artifact, with session logs as its source. It does not require an adopted guideline edit or establish improved capacity. Its qualification and the unresolved alternative routes were independently checked in profile verification; its underlying warrants are MEM-RTE-feedback-review and MEM-OBJ-review-watermarks.
A separate, more ambitious pathway turns retained defects and human reactions into proposals to change standing guidance, checks or hooks. Miners prioritize diagnostic context; the Editor-in-Chief interprets it and proposes edits. The adoption protocol requires operator approval, applicable measurement and later grounding in revised guidance. Full transition rationales persist, but miner summaries truncate them and local rejected-direction notes are unavailable. Schema admission of an accept record does not prove measurement or approval occurred. See MEM-RTE-failure-review, MEM-OBJ-defect-log, MEM-OBJ-local-memory and RT-RTE-guideline-change.
This supports improvement-directed, operator-mediated self-change as a specified pathway, not exercised self-improvement or demonstrated learning in the sense of improved capacity for future action. Observed installation/consumption and subsequent behavior causally depending on a particular change remain untraced. The accepted account does not independently establish all four theory-builder conditions or an instance where criticism of a consumed theory improved future capacity. Reflection is not established merely because the rules concern the system's own work. Autonomous adoption is not the declared mode: the operator retains adoption authority. These are separate conclusions, not a single negative grade, and missing evidence does not establish absent theory criticism or reflection. See RT-RTE-guideline-change and the epistemic bounded conclusion.
The shipped judge experiment adds sampled review judgments and scores, including criticism of lost conditions and hedges. It does not establish production acceptance, independent world truth, or a reviewer-family effect. Its model-defined reference, hidden scorer/question wording and unavailable originals limit replication and attribution. See EPI-RTE-judge-measurement.
For source ingestion and later evidence-aware reuse, the supported architectural contribution is therefore explicit grounding, separated editorial responsibilities, retained provenance and deterministic navigation. For scenarios requiring universal pre-use acceptance, synchronized withdrawal or demonstrated autonomous quality improvement, this evidence cannot establish the necessary guarantees. A traced production candidate with original evidence, delivered review inputs, disposition and later use would strengthen the acceptance account. Enforced admission across all material consumption paths would change the quarantine assessment. For improvement, a specific evidence-shaped update followed through live consumption and causally dependent later operation would change the exercised-pathway assessment; successful improvement would additionally require outcome evidence against the independently specified objective. The report verification and profile verification retain no blockers or declared limits, but do not supply those missing observations.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Private originals and page-production provenance are unavailable. | SRC-1; MEM-OBJ-source, EPI-RTE-editorial-disposition | Shipped source contracts, sampled retained pages and publication procedure | Corpus-wide extraction fidelity, source truth, and a demonstrated candidate-to-acceptance-to-integration chain | Original documents and linked authoring, review, disposition and later-use artifacts for named candidates |
| Host execution and opaque retained state are not inspected. | SRC-1; RT-RTE-role-delegation, RT-RTE-faction-authoring, RT-CMP-qmd, MEM-OBJ-local-memory | Repository role/host contracts and wrappers, not host/provider internals or private memory | Deployed isolation, compaction, complete memory enumeration, actual prompt compliance or retrieval quality | Versioned host/search implementation and scoped context, persistence and invocation evidence |
| Review is conditional and does not quarantine working files. | SRC-1; RT-RTE-source-ingest, RT-RTE-wiki-export, RT-BAP-hook-feedback | Ingest, legacy/direct paths and scoped tool hooks | Universal review-before-write/use or filesystem capability isolation; actual pre-review deployment exposure is also unobserved | Admission enforcement covering every relevant write/consumer route plus deployment traces |
| Lint evidence has bounded semantic reach and delivery. | SRC-1; EPI-RTE-citation-check, EPI-RTE-join-surface, EPI-RTE-desk-check | Citation regexes, targeted versus corpus synthesis output and Desk protocol | Truth from schema PASS, complete join coverage or actual full-span reviewer comparison | Linked full claims, source spans and reviewer reads/verdicts; detector/channel changes if complete automated surfacing is claimed |
| Semantic novelty and warrant of generated narratives remain indeterminate. | SRC-1; EPI-RTE-narrative-authoring, EPI-RTE-editorial-disposition | Query/save, trail/timeline and synthesis procedures without traced premise/candidate pairs | Entailed derivation, ampliative conjecture or accepted new knowledge inferred from a narrative label | Proposition-level input/output comparison and evidence-consuming acceptance against named criteria |
| Freshness and withdrawal are consumer-specific. | SRC-1; MEM-OBJ-dependencies, MEM-RTE-hub-withdrawal, EPI-RTE-reground-surface | Date-based advisories, demotion triage and superseded/open checking | Automatic regrounding, dispute closure, universal archive exclusion or removal from all later reads | Completed maintenance decisions and downstream selection/invalidation evidence across relevant consumers |
| Distribution omits evidence and has separate refresh horizons. | SRC-1; RT-RTE-wiki-export, MEM-RTE-browser-read, MEM-RTE-briefing-read | Export/body-bundle producers, browser loading and email transformations | Complete grounding delivery, installed hosted context membership, synchronized remote refresh or preserved links everywhere | Actual generated/deployed payloads, hosted installation/read evidence and channel-specific refresh or withdrawal controls |
| Diagnostic recurrence and checkpoint adaptation do not demonstrate improved capacity. | SRC-1; MEM-RTE-feedback-review, MEM-RTE-failure-review, RT-RTE-guideline-change | Filtered/clipped feedback, mixed defect histories, operator-mediated adoption and later-read protocol | Causal learning benefit, exercised self-improvement, autonomous adoption or complete trace-learning inventory | Provenanced update/adoption and subsequent causal-use evidence; controlled outcome comparison for benefit; resolved local/host alternatives for inventory closure |
| Experiment artifacts support a bounded model-reference evaluation. | SRC-1; EPI-RTE-judge-measurement, RT-CMP-fable | Sampled labels/scores and experiment plan; no target execution or independent recomputation | Full replication, world-truth validation, production efficacy or isolated reviewer-family effect | Published scorer/questions and originals, full reproducible evaluation and independently varied component comparisons |
No frozen source files were read for this synthesis. Supporting passages come from the supplied accepted records. The reconciliation reports no unresolved inter-report conflict, and both supplied verifications declare Limits none; the rows above preserve the account's evidence boundaries rather than adding new verification findings.