Two-Step Ingest Pipeline

Single-pass ingestion is tempting, but it makes the model read, decide, edit, and audit in one breath. A two-step ingest pipeline separates analysis from writes so source interpretation, contradictions, and update targets can be reviewed before the wiki changes.

hashanalyzestagereviewwritelint

Pipeline stages

Stage	Purpose	Output	Stop condition
Hash and register	Detect duplicate or changed source files.	Path, checksum, source type, intake date.	Unknown origin, unsafe file, or duplicate already reviewed.
Analyze	Read without writing durable pages.	Summary, claims, entities, contradictions, update targets.	Source is unreadable, unsupported, or outside scope.
Stage	Propose exactly which wiki pages should change.	Patch plan or staged draft page.	Target page cannot be identified safely.
Review	Human or maintainer checks risky claims and scope.	Approved, deferred, narrowed, or blocked decision.	Unsupported public claim or authority drift.
Write	Update source summary, concepts, index, and log.	Changed files and final source trail.	Lint fails or contradictions were hidden.

Analysis blueprint

source_path:
source_hash:
source_type:
summary:
claims:
  - claim:
    source_trace:
    status: source-linked | source-needed | contradicted
entities:
contradictions:
update_targets:
review_questions:
do_not_publish:

Staging decision table

Finding	Disposition	Why
Directly source-backed claim that improves an existing page.	Stage an update to the target page.	The graph gets better without creating duplicate pages.
Claim is useful but source evidence is weak.	Stage as source-needed or open question.	The idea remains visible without becoming trusted guidance.
Claim conflicts with current page.	Stage a contradiction record first.	Conflict needs scope/date review before synthesis.
Source is outside scope or unsafe to publish.	Defer or block; do not write public pages.	Intake acceptance is not publication approval.

Write checklist

Create or update one source-summary page before changing concept pages.
Preserve contradictions instead of resolving them by tone or majority vote.
Update wiki/index.md for every new or materially changed page.
Append wiki/log.md with source path, hash, changed pages, and unresolved questions.
Run lint before marking the ingest complete.

What this does not claim

This page is a runbook, not a live ingestion product. LlmWikis does not currently provide automated arXiv ingestion, public MCP access, hosted file processing, or anonymous public editing.