Instructions:Maintenance/Source Integration Review

From Encyclopedia Ephemera
Revision as of 14:16, 12 May 2026 by EphemeraAdmin (talk | contribs) (1 revision imported)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Instruction Metadata
id source-integration-review
type maintenance
applies_to Any
task_type source_integration_review
priority high
status active
canonical true
include_by_default no
requires Instructions:World Bible,Instructions:Core/Canon Policy,Instructions:Core/Continuity Rules,Instructions:Create/Source/Ingest,Instructions:Context/Assembly
tags maintenance,review,integration,source,candidate,queue


Purpose

Given a newly ingested Sources: page and a list of candidate encyclopedia articles identified by the deterministic candidate discovery system, decide what action each candidate requires.

This workflow produces decisions and optionally creates Project: queue tasks. It does not edit encyclopedia articles directly. Article edits are performed by the batch processor (Instructions:Maintenance/Process Pending Source Integrations) after human or agent review of the task queue.

When This Workflow Runs

  • After a Sources: page is created via the ingestion workflow
  • When a human clicks "Run LLM review" on a candidate in the UI
  • When the batch processor invokes it as part of queue processing

Inputs

The context package for this workflow must contain:

The Sources
page
Full content including
Source Metadata
id
type
subtype
author
affiliation
date
location
canonical true
reliability
bias
status published
related
tags
metadata and Content section.
Read the reliability and bias fields before making decisions —
 a low-reliability source should rarely trigger expansion_needed alone.
The candidate encyclopedia article
Full current wikitext, not just the title.
Do not assume from the title alone. Read the article.
The candidate's score and score reasons
Provided by rank_integration_candidates(). Includes which signals
 fired (stub, no_sources, linked_entity, multi_source).
The candidate's existing agent metadata
The HTML comment block at the bottom of the article, if present.
Check last_agent_edit, operation_type, and integration_task_id.
If an open task already exists for this source+candidate pair, do not
 create a duplicate. Update the existing task instead.
Relevant Talk page editorial notes (if include_talk=true in Config)
Bias Analysis and Reliability Assessment sections only.
These may affect the weight given to a low-reliability source.

Read-Before-Write Rule

Before deciding on any action that would create or update a Project: task, re-read the current state of:

  1. The candidate article (may have been edited since the candidate was scored)
  2. The Project:Pending Source Integrations page (check for existing tasks)
  3. The candidate's agent metadata comment

If the article has changed significantly since scoring, re-evaluate. If a task already exists for this pair, merge or update rather than duplicate.

Decision Criteria

For each candidate, output exactly one of the following decisions.

no_action

The article already adequately covers what the source describes. No update is warranted.

Apply when:

  • The specific claims or events in the source are already represented in the article
  • The source adds colour or detail but no new factual content
  • The source is low-reliability and the article already reflects the established view
  • The thematic connection is real but not specific enough to improve the article

This is the expected outcome for most candidates. A review session that creates tasks for more than 30-40% of candidates indicates the confidence threshold is set too low or the scoring is noisy.

citation_only

The source supports a claim already present in the article but the article lacks an inline citation for it. Add a ref tag. No content change.

Apply when:

  • A specific claim in the article is supported by this source
  • The claim is already correctly stated
  • The article would benefit from the citation for verifiability

Do not apply if the source is low-reliability and the claim is contested. A citation from a biased source requires a note, not a bare ref tag.

Output: create a Project: task with action=citation_only, priority=low.

expansion_needed

The source contains substantive new information not represented in the article. The article should be expanded using this source.

Apply when:

  • The source describes events, details, or context the article lacks
  • The information is material — it changes the understanding of the subject
  • The source reliability is medium or higher, or multiple sources converge
  • The article is a stub and this source could meaningfully develop it

Do not apply if:

  • The source is the sole low-reliability source for the claim
  • The information is speculative or propagandistic without corroboration
  • The article would require a major rewrite rather than an expansion

Output: create a Project: task with action=expansion_needed, priority based on score (high if score >= 6, medium if score 4-5).

citation_with_note

The source is relevant but has significant reliability or bias concerns that must be flagged alongside any citation.

Apply when:

  • The source supports or expands the article but is marked low-reliability
 or high-bias in its 
Source Metadata
id
type
subtype
author
affiliation
date
location
canonical true
reliability
bias
status published
related
tags
metadata
  • The information is worth citing but the citation needs a qualifier
  • Example: "According to a MercuryLink corporate advertisement..."

Output: create a Project: task with action=citation_with_note, priority=low. Include the reliability and bias notes from the source metadata in the task reason field.

contradiction_review

The source directly contradicts a specific claim in the existing article.

Apply when:

  • A factual assertion in the source conflicts with a factual assertion in the article
  • The conflict is specific and substantive, not merely a matter of perspective
  • The conflict cannot be resolved by noting both perspectives inline

Do not silently resolve contradictions. Do not choose one version. Do not rewrite the article to match the source.

Output: create a Project: task with action=contradiction_review, status=needs_human. Specify the exact conflicting claims in the task reason field — quote both the article text and the source text precisely.

new_page

The source names or describes a specific entity that has no encyclopedia page and is notable enough to warrant one.

Apply when:

  • The entity is named and specific (a person, place, organisation, event, technology)
  • The entity appears to be significant within the world — not a passing mention
  • Sufficient detail exists in the source (or the existing wiki) to write a stub

Do not apply for:

  • Generic concepts that do not need their own page
  • Entities mentioned only once in passing with no detail
  • Entities that are already red links in many articles (those are handled by
 the stub generation workflow, not integration review)

Output: create a Project: task with action=new_page, priority=medium. Include the source page and the relevant passage as context.

defer

The candidate is related to the source but the decision requires information not currently in the context package.

Apply when:

  • The article references other articles that should be read first
  • The source covers a topic where the wiki's canonical position is unclear
  • The decision requires human domain knowledge not encoded in instructions

Output: create a Project: task with action=defer, status=deferred. Note specifically what additional context is needed.

none

The candidate scored above the threshold but on close reading has no meaningful relationship to this source.

Apply when:

  • The connection is purely thematic with no specific shared content
  • The scoring signal was a false positive (e.g. a common word match)
  • The article and source genuinely do not interact

Output: no task. Log the dismissal in the provenance metadata.

Output Format

Return one block per candidate in this structure:

Target: [[Article Title]]
Score: N  Signals: [stub, no_sources, linked_entity, ...]
Decision: no_action | citation_only | citation_with_note | expansion_needed |
          contradiction_review | new_page | defer | none
Reason: One or two sentences. Be specific — cite the source claim and the
        article state that led to this decision.
Task created: yes / no
Task action: [action value if yes]
Task priority: [high / medium / low if yes]

After all candidates, add a summary line:

Summary: N candidates reviewed. N tasks created (N expansion_needed, N citation_only, N citation_with_note, N contradiction_review,
N new_page, N defer). N dismissed (no_action / none).

Constraints

  • Read the full article text before deciding. Never assume from the title.
  • Read the source reliability and bias fields. Weight decisions accordingly.
  • Most candidates should receive no_action or none. If the majority receive
 expansion_needed, the threshold is too low.
  • A single source should not generate more than 5 tasks. If more seem warranted,
 create the 5 highest-confidence ones and note the rest as deferred.
  • Never silently resolve contradictions. Always escalate to needs_human.
  • Never create duplicate tasks. Check existing tasks before creating a new one.
  • The defer decision is not a failure. Use it when the decision requires
 information this workflow does not have.

Reliability Weighting

Use the source

Source Metadata
id
type
subtype
author
affiliation
date
location
canonical true
reliability
bias
status published
related
tags
metadata to weight decisions:
Source reliability Effect on decision
high Full weight. Apply decisions normally.
medium Apply normally but note reliability in task reason for expansion/contradiction cases.
low Do not trigger expansion_needed from a single low-reliability source alone.

Apply citation_with_note. Escalate contradiction_review to needs_human regardless.

unknown Treat as low. Flag for reliability assessment in the task.

Multi-source convergence overrides single-source reliability constraints. If three low-reliability sources independently describe the same event, that convergence is itself a signal worth noting.

Provenance

After completing the review, the calling system logs the following to the source page's agent metadata:

{
  "operation_type": "source_integration_review",
  "candidates_reviewed": N,
  "tasks_created": N,
  "decisions": {
    "no_action": N,
    "citation_only": N,
    "citation_with_note": N,
    "expansion_needed": N,
    "contradiction_review": N,
    "new_page": N,
    "defer": N,
    "none": N
  }
}