LLM Ledger

by Verbio Labs
A verifiable, time-aware knowledge ledger your LLM keeps for you.

English · 한국어

A permanent markdown knowledge base that an LLM synthesizes and maintains for you —
an evolution of Karpathy's "LLM Wiki" pattern,
implemented on top of Claude Code.

But instead of accumulating prose pages like a wiki, it accumulates atomic "claims" —
each with its sources, a confidence level, and a validity window in time. You do the sourcing
and the asking; the LLM does every write, every verification, and every cross-reference.

No external skills or plugins. This folder alone runs in any Claude Code environment.
Take it, fill it with your own data, and grow it as you go — a starter, not a final answer.

Why a "ledger," not a "wiki"

The original LLM Wiki's atomic unit was a synthesized prose page. Powerful — but three things were structurally impossible:

In one line: a wiki is mutable prose; a ledger is append-only and auditable.
We don't overwrite facts — we supersede them and keep the past.

And it inherits the original's best idea — constant token cost. index.md is not a catalog;
it's a router (MOC) that takes the intent of a question and decides which shard to open. Claims are
indexed by topic, time, and confidence, so the per-question token budget stays flat as the ledger grows.

The difference at a glance (shipped as a live example)

"As of 2005, was Pluto a planet?"

• A normal wiki: "Pluto is a dwarf planet" — overwritten by the 2006 fact, so it gives the wrong answer.
• LLM Ledger: the old claim is preserved as superseded, so it answers "Yes — in 2005 it was a planet." — the correct as-of answer.

See it in the repo: 30-ledger/claims/pluto.md · 50-queries/2026-06-19-pluto-as-of-2005.md

Quick start

cd llm-ledger
claude

A usage guide pops up automatically on first run (a SessionStart hook).

1. Add a source — /ingest <URL · file · text> saves the original into 10-inbox/.
2. Ledger it — /compile extracts claims from the source, assigns provenance / confidence / validity,
   resolves conflicts (supersede or contested), refreshes topic views and indexes, then moves the original to 20-raw/.
3. Ask — /query Was Pluto a planet in 2005? retrieves and cites claims via two-stage routing.
4. Check — /audit scans for contradictions, low confidence, and temporal inconsistency.
5. Trace — /timeline pluto shows how the knowledge changed over time.

In inbox = not yet compiled. In raw = compiled.

Architecture (4 layers)

10-inbox/ → /ingest → /compile → 30-ledger/
                                  ├── claims/      (atomic facts — the source of truth)
                                  ├── topics/      (read-only views assembled from claims)
                                  ├── index.md     (router MOC)
                                  ├── indexes/     (by-topic / by-time / by-confidence)
                                  └── aliases.md   (canonicalization = routing keys)
                      /query    → Phase A routing → Phase B retrieve & cite → 50-queries/ file-back
                      /audit    → contradictions · low-confidence · orphans · temporal checks
                      /timeline → unfold claims chronologically; reconstruct as-of snapshots
                      originals move to 20-raw/ after processing (kept immutable)

A claim — the atomic unit

id: clm-2026-0002
statement: "Pluto is the ninth planet of the Solar System."
sources:
  - ref: 20-raw/pluto-tombaugh-1930.md
    quote: "it was announced as the ninth planet of the Solar System"
confidence: high
valid_from: 1930-02-18
valid_until: 2006-08-24      # ← temporal boundary
status: superseded            # ← preserved, not overwritten
superseded_by: clm-2026-0003

A topic page is a view assembled from claims like these, with a [^clm-id] footnote on every
assertion. If a view is ever damaged, it can be regenerated from the claims at any time.

Commands

Validation (trust, not vibes)

The ledger's correctness shouldn't depend on the LLM behaving. A tiny zero-dependency
validator enforces it:

python3 tools/ledger.py check     # exit 1 if anything is wrong
python3 tools/ledger.py search "your question" --as-of 2022-12-31
python3 tools/ledger.py stats

check verifies every claim has provenance, valid enums, a real source file, sane
temporal bounds (valid_from <= valid_until), and reciprocal supersession links —
plus footnote and index integrity across topic views. It runs on every push via
GitHub Actions, so a broken ledger fails CI instead of silently rotting.

Learn more

• Operating contract: CLAUDE.md
• Validator & CLI: tools/ledger.py
• Full spec (claim schema, confidence, conflict handling, routing, sharding, provenance): 00-system/conventions.md
• Templates: 40-templates/

Lineage & credits

Inspired by fivetaku/llm-wiki and
Karpathy's LLM Wiki gist,
extending the concept toward "claims as the unit + provenance/confidence/time + conflicts preserved."

MIT License. Fork it freely.

_{🦉 Verbio Labs}