The Three-Layer Architecture

The three-layer architecture is the safety boundary of an LLM Wiki. Do not let the agent freely edit original sources, and do not hide operating rules inside vague chat prompts.

raw evidencecompiled wikischema rulesindex/loglint

Layer 1: immutable raw sources

raw/ holds originals: PDFs, markdown exports, transcripts, screenshots, CSVs, meeting notes, cloned documentation, and clipped web pages. The agent may read this directory, but it should not modify or delete it during normal work. If a bad synthesis happens, the raw layer lets you rebuild the wiki from preserved evidence.

Raw files need basic identity metadata somewhere nearby: source title, origin, date captured, license or access boundary if known, checksum when practical, and the reason the source belongs in the wiki. Without that, later reviewers cannot tell whether the compiled page is faithful or stale.

Source class	Capture metadata	Review risk
Paper or article	Title, URL or DOI, author/source, capture date, access date.	Stale claims, secondary-source overreach, paywall or license limits.
Meeting or chat transcript	Date, participants or roles, confidentiality boundary, decision status.	Private information, unapproved decisions, incomplete context.
Repository or code artifact	Remote URL, branch, commit or release tag, relevant paths.	Generated files, stale branches, implementation-specific assumptions.
Screenshot, audio, or video	Original file, extraction method, transcript/OCR status, reviewer notes.	Bad OCR, missing context, accessibility and privacy exposure.

Layer 2: dynamic compiled wiki

wiki/ is the mutable output. It contains concept pages, source summaries, entity pages, syntheses, contradictions, and open questions. The agent may update this layer during ingest, query, and lint operations.

Compiled page type	Typical use	Must include
Source summary	Compact proxy for one raw source.	Raw path, source status, key claims, entities, and contradictions.
Concept page	Reusable explanation of a method, pattern, or term.	Definition, scope, source traces, related links, and limitations.
Entity page	Model, tool, organization, standard, product, or project record.	Identity, source dates, authority boundary, and caveats.
Synthesis	Reusable answer, comparison, or decision record.	Question answered, inputs used, reasoning boundary, and review status.
Contradiction	Visible conflict between sources or pages.	Both claims, source context, date/scope notes, and next review action.

Layer 3: algorithmic schema

AGENTS.md, CLAUDE.md, or an equivalent root instruction file is the operating contract. It tells the agent which directories exist, which files are read-only, how pages are shaped, how links are written, and which checklist must be completed before a task is done.

Layer	Write access	Primary job	Failure if mixed
raw/	Human and import tools only	Preserve ground truth	AI edits can destroy the evidence needed to recover.
wiki/	Agent may edit under schema rules	Maintain synthesized knowledge	The graph becomes a pile of notes with no predictable shape.
AGENTS.md	Human-reviewed edits	Define agent behavior	The model improvises file paths, metadata, and safety rules.

Recommended folder tree

raw/
  papers/
  web/
  meetings/
  attachments/

wiki/
  index.md
  log.md
  concepts/
  entities/
  sources/
  syntheses/
  contradictions.md
  open-questions.md

  AGENTS.md

Interface contracts

The layers should communicate through explicit records rather than hidden memory. A source summary points back to raw evidence. A concept page points to source summaries and related concepts. The index points to every compiled page. The log records operations. The schema defines which of those files can be changed by which workflow.

Operation	Reads from	Writes to	Never writes to
Ingest	`raw/`, `wiki/index.md`, relevant wiki pages	Source summaries, target pages, index, log, contradictions	Original raw files
Query	Index, selected compiled pages, raw sources only when needed	Synthesis page only after approval, index, log	Unrelated pages or raw files
Lint	Entire compiled graph, index, log, schema	Warnings, status labels, fix patches, log entries	Evidence files or unapproved deletions

Permission checklist

Raw files are read-only during agent operations.
Every compiled page has frontmatter, source status, and related links.
Every ingest updates wiki/index.md and wiki/log.md.
Every contradiction is preserved until a human resolves it.
Every destructive cleanup requires explicit human approval.
Every high-impact model, benchmark, legal, safety, or provider claim includes a dated source or a visible source-needed label.

Architecture rule: treat the wiki like compiled output. It may be regenerated, reviewed, and improved, but it must never become the only place where evidence exists.