Why NeuralWikis Agents Produced Low-Quality, Repetitive Public-Wiki Posts

Full report

On this page

Executive summary

The public record points to an upstream authoring and orchestration problem, not a publication-layer moderation problem. NeuralWikis’ public wiki contribution schema allows unauthenticated posting to the public wiki (auth: "none", publicWikiMutationAllowed: true), but it also states that the optional OpenAI call is moderation-only and does not generate or rewrite article content. Recent visible entries likewise report moderation metadata such as provider: "openai_moderations" with generatesContent: false and rewritesContent: false. That means the low-quality prose almost certainly came from the agents or prompts that authored the submissions before publication, not from the publish API itself.

The strongest observed failure modes are repetition by template, title combinatorics, weak grounding, excessive boilerplate, duplicate and near-duplicate publication, and upstream output leakage. On June 13, 2026, the public wiki listed 595 readable articles across ten categories, while also noting 649 archive files accounted for and 59 hidden duplicate/quarantined/merged records. Within that corpus, large clusters of near-identical titles and summaries were published seconds apart, often under source labels such as Autonomous lesson intake live-deterministic-safety-only-100-v7 and Autonomous lesson intake live-deterministic-safety-only-final-v8, with contributor label NeuroWikis autonomous lesson intake.

Representative examples show that the system was optimizing for “safe transformation” and boundary disclaimers rather than distinct, evidence-rich teaching. Many articles use the same section scaffold, the same abstract phrasing, and the same public-boundary boilerplate, regardless of topic. More severe failures include duplicated entries titled “Teleodynamic as the Ecosystem Fulcrum”, unresolved cross-environment citation tokens such as turn1view0, repeated paragraphs, and a truncated body ending in | Dimensi..., all of which indicate missing sanitation and output-completeness checks before publication.

The most likely root cause is a multi-stage “lesson intake” pipeline that converts bounded source digests into many public articles by permuting topic keywords through a rigid template, then passes the result through safety moderation without semantic deduplication, citation enforcement, markdown completeness checks, or editorial scoring. Public source-intake pages reinforce that the system is heavily centered on constrained summaries, excluded claims, quarantine, and mutation boundaries; what is missing from the public evidence is an equally explicit quality gate for uniqueness, specificity, or evidence density.

The highest-priority fix is to redesign the pipeline around grounded extraction, constrained drafting, and automated verification before publish. In practice, that means: generate from evidence tables rather than keyword bundles; require article-specific citations; run semantic deduplication against recent posts; reject unresolved citation markers and incomplete markdown; keep extraction and verification passes near-deterministic while allowing only slight variability in final prose; and promote only articles that clear measurable QA thresholds. These recommendations align with public prompt-engineering guidance that emphasizes specificity, explicit output formats, “when unsure” fallbacks, grounding data, tool/input examples, and test-and-iterate evaluation.

Scope and method

This report is based only on public evidence available on the NeuralWikis and NeuroWikis sites, plus public prompt-engineering and retrieval literature. I assumed no access to internal logs, unpublished prompts, unpublished repositories, or private reviewer records. That assumption is reinforced by the public contribution schema and entry metadata, which say raw contribution payloads are not stored or echoed on the public surface.

The main sources used were the public wiki index and representative entry pages on neurowikis.com, the public contribution API schema and entry metadata on neuralwikis.com, the source-intake review queue, llms.txt, the trust policy, the agent quickstart, and the connection text intended for chatbots. Those public agent-facing materials are important because they are the closest visible equivalents to system-level guidance, even though they are not the exact private prompts used by the posting agents.

I sampled representative posts from three visibly different publishing patterns: the older Codex archive-verifier pattern, the more readable but still generic live-openai pattern, and the highly repetitive live-deterministic-safety-only pattern. That mix is enough to support a high-confidence diagnosis of the failure modes, but it cannot prove the exact private prompt strings, temperature values, seeds, or loop logic used inside the unpublished orchestration layer. Where those details are not directly visible, I state them as inferences.

Representative sampled posts

What the public record shows

The public wiki explicitly frames itself as a surface for “scanner-clean public contributions” that appear as transformed teaching articles rather than canonical KB records. The index reports 595 readable public articles, 649 archive files accounted for, and 59 quarantined/duplicate/merged records hidden from public reading, spread across ten categories. That means some duplicate handling exists, but the visible corpus still contains obvious duplication and templated saturation.

The contribution API schema is unusually revealing. It says the article list and create/read endpoints are public, the contribution surface uses auth: "none", the public wiki store is writable, canonical KB mutation is not allowed, and the optional OpenAI safety review is moderation-only via /v1/moderations, with generatesContent: false and rewritesSubmittedContent: false. The same theme appears in entry metadata for published items, which show moderation flags but explicitly say the public boundary does not generate content with OpenAI and does not rewrite user wiki input with OpenAI.

The source-intake layer is safety-rich but editorially thin. Its public review queue shows 29 visible source reports represented as public-safe digests, with explicit excluded claims and dispositions such as accepted_for_constrained_seed_use or deferred_roadmap_only. The surrounding language is consistently about bounded summaries, excluded claims, quarantine, and mutation being false. That posture is coherent for safety, but it also strongly suggests an upstream pipeline that compresses source material into safe abstractions before generation, which makes specificity loss very likely if there is no second-stage grounding pass.

The agent-facing guidance surfaces reinforce that diagnosis. llms.txt tells agents to read safety and routing guidance first, use semantic search before broad crawling, avoid blind imports, and keep source promotion behind protected routes. The trust policy says external input starts outside active memory, prompt-injection-like input is treated as hostile, and excessive agency is denied unless bounded. The quickstart tells agents to inspect public routes first and stop if the task is actually a human explanation request. The chatbot connection text is “read-only and citation-first,” explicitly telling agents to cite canonicalUrl, trustCategory, and citations from the public KB. All of this public guidance is excellent for safety and routing, but none of it defines a strong editorial objective for writing distinct, source-rich public articles.

The publication cadence is also a major clue. The index shows dense bursts of posts published within seconds, including a block around 12:36:02–12:36:05 UTC and another around 13:53–14:05 UTC, many of them sharing the same title stem and only rotating keyword triplets such as Page / Audit / Evidence, Page / SEO / Audit, Agent / JSON / API, or Semantic / Schema / Constraint. That pattern is much more consistent with batch keyword permutation than with human curation or article-by-article synthesis.

The following diagram summarizes the most likely current public pipeline, inferred from the site’s public docs, metadata, and visible outputs. The key point is that public evidence shows safety gating and moderation, but not quality gating.

flowchart LR
    A[Source reports and archived markdown] --> B[Bounded source digest intake]
    B --> C[Autonomous lesson intake agent]
    C --> D[Keyword extraction and title permutation]
    D --> E[Template-based article drafting]
    E --> F[Moderation-only safety screen]
    F --> G[Public wiki publish]

    E --> H[No semantic dedup]
    E --> I[No evidence-density gate]
    E --> J[No citation-marker sanitizer]
    E --> K[No markdown completeness validator]

Why the posts degraded

The most obvious failure is template collapse. Many articles share the same skeleton: a title stem plus keyword triplet, a short “transformed teaching article” summary, then the same sequence of sections—What This Teaches, Reusable Pattern, Common Mistakes, Safe Application Checklist, Where This Fits in NeuroWikis / NeuralWikis, and Public Boundary. In the deterministic batch, that sameness becomes extreme enough that changing the topic mostly changes only the inserted keywords. This is especially visible in the “Keeping AI Agent Workflows Bounded and Auditable,” “Turning Site Reviews Into Safer Public Guidance,” and “Keeping AI Research Claims Proportional to Evidence” clusters.

A second failure is semantic near-duplication. “Using Immutable Governance Anchors for Agent Safety” and “Designing Immutable Policy Anchors for Agent Governance” are not exact copies, but they cover the same core pattern with highly overlapping structure, examples, and warnings. The public wiki also shows two separate entries with the exact same title, “Teleodynamic as the Ecosystem Fulcrum.” That visible duplicate pair matters because the index already claims that some duplicates are hidden; if these duplicates still survived, the deduplication logic is either too shallow or too literal.

A third failure is weak grounding and evidence drift. Some posts are titled as if they will discuss evidence, research claims, or specific site reviews, but the actual bodies are generic, citation-free abstractions. The best example is the “Keeping AI Research Claims Proportional to Evidence” article: it speaks in generalities about bounded claims and uncertainty, yet provides no concrete evidence excerpts, no external citations, and no site-specific comparisons. That stands in tension with the site’s own public connector guidance, which tells agents to use bounded context retrieval and cite canonical URLs and supporting citations.

A fourth failure is boilerplate overweighting. Even the more readable live-openai-* articles spend a large share of their token budget on repeated disclaimers that the article does not claim private workspace access, billing, source promotion, adoption approval, rollback execution, or operator decisions. Those boundaries are reasonable, but repeating them across nearly every article crowds out distinctive substance. In practice, the public-boundary section became a mandatory ritual rather than a proportional disclaimer.

A fifth failure is upstream artifact leakage and incomplete outputs. The “Teleodynamic as the Ecosystem Fulcrum” entries contain unresolved turn... citation tokens that only make sense inside another AI product’s rendering environment, duplicate their opening paragraph, and terminate in a truncated table fragment. Because the public contribution API and the newer deterministic entries both indicate that publication-layer OpenAI use is moderation-only and not rewriting content, these defects almost certainly originated in the upstream authoring agent or a copy-through archival/import path.

Inferred prompt and orchestration pattern

The exact private prompt is not publicly visible, but the repeated output structure supports a high-confidence reconstruction of the likely template family:

You are NeuroWikis autonomous lesson intake.
Transform an archived source into a non-verbatim public lesson.

Requirements:
- Produce a public-safe teaching article.
- Use title family: <prefix>: <kw1> / <kw2> / <kw3>
- Include sections:
  What This Teaches
  Reusable Pattern
  Common Mistakes
  Safe Application Checklist
  Where This Fits in NeuroWikis / NeuralWikis
  Public Boundary
- Emphasize excluded claims and protected behaviors.
- Do not expose raw source text or private implementation details.

That inference is supported by the repeated section layout, repeated “transformed public lesson” summaries, repeated keyword-slot titles, and repeated public-boundary language across otherwise unrelated topics.

The most likely orchestration pattern is: source digest → keyword extraction → one-to-many template expansion → moderation-only screen → immediate publication. The source-intake queue’s “accepted constrained seed use” language and the public source labels like live-deterministic-safety-only-100-v7 and final-v8 fit that model well. So do the bursty timestamps and combinatorial title families.

Parameter and settings assessment

Some settings can be inferred with moderate confidence. The live-deterministic-safety-only-* source labels, together with the visible sameness of wording, strongly suggest a low-variance or seeded deterministic generation mode for those batches. Microsoft’s public guidance notes that lower temperatures make outputs more focused and concrete, while higher temperatures increase divergence, and its reproducible-output documentation notes that baseline model behavior is otherwise nondeterministic. That does not prove an exact temperature value, but it does support the conclusion that these batches were likely authored under intentionally low variability.

However, temperature is probably a secondary cause. The main cause of repetition is the apparent looped prompting over the same template with different keyword bundles, combined with missing deduplication. Raising temperature alone would only make the repetitive system slightly noisier; it would not solve the fact that the content plan itself appears repetitive.

The Teleodynamic failures also suggest one or both of: insufficient max-token budget, and no completion validator/stop-condition validator. The body ends mid-thought and mid-table, which is exactly the kind of defect that should be caught by a markdown-completeness or required-sections check. Public prompt guidance from Microsoft and Anthropic emphasizes explicit output formats, clear syntax, examples, and testing/iteration; those protections are not visible in the published artifact stream.

The evidence also supports a context-management problem. Source-intake digests are bounded and safety-heavy, but the resulting articles often lose concrete evidence and collapse into generic reusable patterns. That is consistent with a system that compresses long source material too early, then drafts from summaries or keyword bundles rather than from a structured evidence table. Long-context research supports this concern: “Lost in the Middle” shows that long-context performance degrades when relevant information sits in the middle of the input, while RAG literature shows that retrieval-augmented generation improves specificity and factuality versus relying only on parametric recall.

Recommended remediation

The most effective fix is to stop treating “safe transformation” as the only gate. The current public surfaces show strong safety boundaries, but high-quality public teaching requires a second set of gates for grounding, uniqueness, evidence density, completeness, and editorial fit. Public prompt-engineering guidance supports that approach: define success criteria first, specify the job and boundaries clearly, enforce output format, give the model a “when unsure” policy, ground on explicit source material, and test with realistic and adversarial cases. Tool guidance likewise recommends detailed descriptions and schema-validated examples for complex tool behavior.

Priority redesign

Use a four-stage pipeline instead of single-pass posting:

Input: source digest plus retrievable supporting passages. Output: structured JSON with claim, evidence_snippet, source_url, confidence, excluded_claims, article_angle. Recommendation: temperature 0.0–0.2; fixed schema; no prose generation. This stage should reject the task if fewer than three concrete supporting facts are available. This follows site guidance to use structured routes and grounded context, not broad freeform drafting.

1. Evidence extraction pass

Input: validated evidence JSON plus recent-title/body fingerprints. Output: one angle only, with a novelty explanation such as “how this differs from the closest existing post.” Recommendation: semantic dedup before drafting. If similarity to an existing post exceeds a threshold, return MERGE_CANDIDATE instead of drafting. The visible duplicate and near-duplicate failures make this step urgent.

1. Distinctive outline pass

Input: outline plus evidence table only. Output: final article markdown with short boundary note and inline citations. Recommendation: temperature 0.3–0.5, top_p=1, alter only one randomness parameter, and keep the system message short enough that it does not crowd out the evidence context. Microsoft guidance explicitly warns that longer system messages consume context and that ordering matters; evidence should sit close to the generation step.

1. Grounded article pass

Output: PASS, FAIL, or HUMAN_REVIEW. Checks: no unresolved citation tokens, no duplicated paragraphs, no truncated tables/sentences, minimum evidence count, semantic dedup, boilerplate ratio, and article-type policy match. This is where operational tests such as the safety-review retest should be filtered out of the educational public surface.

1. Verifier pass

Recommended settings and controls

The table below maps the current public failures to concrete fixes and expected impact.

Context-window and chunking strategy

For long or messy source material, do not draft directly from a full report or a long conversation transcript. Public Microsoft guidance stresses space efficiency and grounding, Azure’s “On Your Data” flow explicitly separates ingestion, retrieval, filtration, reranking, and document inclusion, and long-context literature shows that relevant information in the middle of a large prompt is easy to miss. A safer design is structural chunking plus hierarchical summary: section-based chunks in the few-hundred-token range, light overlap, retrieval/reranking over those chunks, then a compact evidence table for drafting. The exact chunk sizes are engineering recommendations rather than facts surfaced by the site, but this general architecture is well supported.

A practical implementation would keep source chunks and evidence snippets separated from prose instructions. The model should see: job, output schema, evidence table, excluded-claims table, nearest-neighbor duplicates, and then draft. It should not see unbounded archive text mixed with unrelated workflow history, because that is the configuration most likely to produce generic summaries, lost-middle omissions, and citation-marker leakage.

Sample corrected system message

You are the Public Wiki Lesson Generator for NeuroWikis.

Your job is to write ONE distinctive, evidence-grounded public article from the supplied evidence table.

Hard rules:
- Use only facts present in the evidence table.
- If evidence is insufficient, output NEEDS_REVIEW instead of drafting.
- Do not output internal tool traces, session ids, citation renderer tokens, or source-control metadata.
- Do not repeat paragraphs or reuse stock boundary boilerplate beyond a single short boundary note.
- If the proposed article is semantically too similar to an existing article, output MERGE_CANDIDATE.
- If markdown is incomplete or any citation is unresolved, output FAIL_VALIDATION.

Required output sections:
- Summary
- Key lesson
- Evidence-backed explanation
- Practical takeaway
- Boundary note

Style:
- Human-readable, specific, concise.
- Prefer concrete comparisons over abstract slogans.
- No keyword-triplet titles.
- No generic “transformed public lesson” language.

This system message follows public best-practice guidance to define the assistant’s job, boundaries, output format, and “when unsure” behavior clearly, while keeping the message short enough to preserve room for grounded evidence.

Sample corrected drafting prompt

<task>
Write one public article for NeuroWikis.
</task>

<article_goal>
Teach a distinctive lesson that is supported by the evidence below and is not redundant with existing public pages.
</article_goal>

<existing_neighbors>
1. "Using Immutable Governance Anchors for Agent Safety"
2. "Designing Immutable Policy Anchors for Agent Governance"
</existing_neighbors>

<novelty_requirement>
Explain in 2 sentences how this article will differ from the nearest existing page.
If you cannot explain the difference clearly, return MERGE_CANDIDATE.
</novelty_requirement>

<evidence_table>
- Claim:
- Evidence snippet:
- Canonical source URL:
- Confidence:
(repeat for each item)
</evidence_table>

<excluded_claims>
List claims that must not be made.
</excluded_claims>

<output_format>
Return valid JSON only:
{
  "status": "PASS|MERGE_CANDIDATE|NEEDS_REVIEW|FAIL_VALIDATION",
  "title": "",
  "summary": "",
  "novelty_rationale": "",
  "body_markdown": "",
  "citations": [{"url": "", "claim": ""}]
}
</output_format>

This prompt deliberately uses a structured output format, a novelty guard, and an explicit fail path. That design is consistent with public guidance on explicit formatting, giving the model an “out,” and using clear syntax for predictable parsing.

Sample verifier prompt

You are the Public Wiki Publication Verifier.

Given a candidate article and recent published articles, return PASS or FAIL.

Fail if any of the following are true:
- unresolved internal citation markers exist
- body ends mid-sentence or with broken markdown
- repeated paragraph similarity > 0.90
- semantic similarity to a recent article > threshold
- fewer than 3 grounded citations support the body
- boundary note exceeds 15% of article tokens
- title follows a keyword-triplet template
- article is an operations test, runtime diagnostic, or deployment retest rather than a public lesson

Return JSON only:
{
  "status": "PASS|FAIL",
  "reasons": [],
  "similar_articles": []
}

Anthropic’s public tool guidance is especially relevant here: detailed descriptions, schema-validated examples, and clear constraints improve agent/tool performance, while Microsoft guidance stresses testing and iteration of prompts against realistic failures.

Action checklist

• Disable direct auto-publish from authoring agents; route all new drafts through a verifier gate first.
• Require a structured evidence table with canonical source URLs before drafting.
• Add semantic deduplication on titles, outlines, and final bodies against recent posts.
• Ban keyword-triplet titles unless a human editor explicitly approves them.
• Split deterministic extraction from lightly variable final prose; do not use one setting for every stage.
• Add sanitation checks for unresolved citation markers, broken markdown, repeated paragraphs, and cut-off endings.
• Convert long repeated boundary paragraphs into one compact injected boundary note.
• Create an eval set from the observed failures: duplicate titles, near-duplicate governance pages, ops-test content, citation-token leakage, and truncated markdown.

Open questions and limitations

The exact private prompts, seeds, temperature values, stop sequences, and orchestration DAG are not publicly exposed, and the public contribution surface says raw payloads are not stored or echoed. Because of that, the prompt templates and parameter settings in this report are necessarily inferred from visible outputs, source labels, and public docs rather than recovered from internal logs.

I also did not identify a publicly accessible repository that exposes the posting-agent code path itself. The report therefore relies primarily on the site’s own public pages, metadata, schemas, and agent-facing guidance, plus established prompt-engineering and retrieval literature. That is sufficient for a high-confidence diagnosis of the observable failure modes and the most likely architectural causes, but not for a byte-level reconstruction of the hidden prompt stack.