Ateneo-KB Public

An editorial knowledge base where AI assists compilation but humans govern judgement.

The problem

Most LLM-assisted knowledge systems optimise for recall: find more, retrieve faster, summarise better. But in domains where interpretation matters — philosophy, philology, psychoanalysis, historiography — the hard problem is not finding a passage. It is deciding how far a reading can travel before it stops being evidence and becomes conjecture.

Current tools collapse that distinction. A retrieved fragment looks the same whether it is a verified quotation or a plausible hallucination. Confidence is invisible. Provenance is implicit. The editorial layer — who decided this claim is supported, at what level, from which source — disappears.

Ateneo-KB is a working method for keeping that layer visible.

Current scope

This public edition is a working sample of the method, not the full research system.

Operational here today:

• immutable raw captures and curated notes with explicit confidence
• generated audit surfaces (graph, lineage map, editorial dashboard)
• reproducible search trails before curation when a source is still being fixed
• hybrid ingest helpers with provenance sidecars and local EPUB cleanup

Not included here:

• the full private corpus
• the live Supabase research bridge used in the main workspace
• the broader internal review and evidence-pack workflow still evolving in the private vault

What this is

A structured, replicable knowledge base that separates:

• Raw captures — immutable source material, never rewritten
• Curated sources — human-reviewed notes with explicit provenance and confidence
• Concepts — interpretive nodes that distinguish firm evidence from open reading from working hypothesis
• Outputs — derived pieces (memos, arguments, public writing) that trace back to curated material

Every note declares its sources, its confidence level (firme · media · abierta), and its epistemic status (evidencia firme · lectura abierta · hipótesis abierta). Nothing enters the curated layer without review. Nothing claims more support than it has.

When a source is not yet fixed, the method also expects a small lookup trace before curation: a raw note that records the queries run, the catalogs or databases checked, the candidate records found, and the identifier or URL that justified the provisional selection.

What this is not

This is not another second brain, not a note-taking app, not an AI-powered personal wiki. The architecture follows Karpathy's LLM Wiki pattern (raw → wiki → outputs), but adds what that pattern intentionally leaves open: editorial governance for domains where getting it wrong has consequences.

Structure

00_inbox/      unprocessed captures (excluded from public sample)
01_raw/        immutable source material
02_wiki/       curated notes, concepts, projects, and generated indices
03_outputs/    derived public-facing pieces
04_ops/        scripts, templates, operational log

Generated surfaces

Three scripts produce auditable views of the knowledge base:

• Wiki graph — dependency map across all notes (wiki_graph.py)
• Lineage map — traces how raw captures become curated claims and then public outputs (render_lineage_map.py)
• Editorial dashboard — review queue, confidence distribution, and proof surface (render_editorial_dashboard.py)

Two short companion docs make the public sample easier to inspect:

• QUICKSTART.md — run the sample in a few minutes
• docs/review-flow.md — see how a raw source becomes a reviewed concept and then an output

Public sample

This repository contains a minimal curated sample — enough to demonstrate the method, not to replicate a full working corpus. The sample includes:

• a denser psychoanalytic chain (Lacan/Miller) that shows editorial judgement under copyright constraints
• a fully public-domain Stoic chain (Epictetus) that can be audited end to end by anyone cloning the repo

Together they show both depth and portability.

Relation to Ateneo

This method powers the editorial research behind Ateneo, a citation verification instrument for humanistic corpora. Ateneo verifies quotations against indexed primary sources; Ateneo-KB governs the knowledge layer underneath — where interpretive claims are tracked, sourced, and kept honest.

Running the scripts

python 04_ops/scripts/wiki_graph.py          # dependency graph
python 04_ops/scripts/wiki_graph.py --strict  # strict mode (flags broken links)
python 04_ops/scripts/render_lineage_map.py   # raw → source → concept → output lineage
python 04_ops/scripts/render_editorial_dashboard.py  # editorial review dashboard

Requires Python 3.10+. No external dependencies beyond the standard library.

If you want the fastest route through the sample, start with QUICKSTART.md.

Ingesting to raw

The public repo also includes two lightweight ingest helpers:

python pdf2md.py path/to/file.pdf
python ingest_to_raw.py path/to/file.docx
python ingest_to_raw.py path/to/file.html --write-meta
python ingest_to_raw.py path/to/book.pdf --converter pdf2md

The intended split is:

• pdf2md.py for PDFs when page-sensitive extraction matters more than broad format coverage
• ingest_to_raw.py as a hybrid front door:
  • pdf -> MarkItDown by default
  • md / txt -> passthrough
  • epub -> local epub_clean extraction to avoid XHTML/CSS noise and missing chapters
  • docx, pptx, xlsx, html, etc. -> MarkItDown if installed

The wrapper now preserves the original source by default unless the input already lives in 01_raw/. Use --skip-original-copy only for explicit exceptions. For PDFs, auto now prefers MarkItDown; use --converter pdf2md when page-sensitive extraction or scanned-page reconstruction matters more than speed. For EPUBs, auto now prefers epub_clean rather than the generic path.

MarkItDown is optional rather than required:

pip install "markitdown[all]"

If you are working on Windows, prefer ingest_to_raw.py over the raw markitdown CLI so the ingest writes files and provenance metadata consistently.

That keeps the sample honest about a core distinction in the method: conversion is useful, but editorial governance still happens later in 02_wiki/.

Search Trails

Use 04_ops/templates/search-trail.md when a source must be located or disambiguated before ingest.

The point is simple: lookup is not yet curation. A search trail should preserve:

• the query you started from
• the databases or catalogs you checked
• the candidate records you found
• the identifier or URL that justified the provisional selection
• the ambiguity that still remains, if any

Replicating the method

The value here is the pattern, not this particular corpus. To adapt Ateneo-KB to another domain:

1. Define your raw capture sources and keep them immutable
2. Use the note templates in 04_ops/templates/ — each enforces provenance and confidence metadata
3. Separate what you can defend from what you are reading into the text
4. Run the scripts to generate auditable surfaces
5. Let the dashboard tell you where your knowledge base is strong and where it is exposed

Design principles

• Criterion over access. Retrieval is solved. Judgement is not.
• Confidence is metadata, not rhetoric. Every note declares firme, media, or abierta — and means it.
• Provenance is structural. Sources link to raw captures. Concepts link to sources. Outputs link to concepts. The chain is always traversable.
• Silence is a valid answer. If there is no evidence, the note says so. The system does not hallucinate support.
• The human governs. AI assists capture and compilation. The editorial decision — what to promote, what to hold, what to flag — stays with the researcher.

Licence

Scripts: MIT Editorial content and templates: CC BY 4.0

A short split-licence summary is also available in LICENSE.

Author

Pablo Martínez Samper — pablomartinezsamper.com