klode

A klode — a rich vein — of grounded, verifiable knowledge, with the machinery to encode it and to supervise work against it. Every claim is anchored to a verbatim source and checked by a fail-closed linter, so a citation can't quietly rot: cite, don't recall.

That guarantee is referential, not semantic — it proves the quoted text is still there, in a current source, in exactly one place. It does not prove the quote supports the claim built on it. Green klode check means "no citation rot," never "this is true."

A second, narrower limit sits beside it. klode ingest verifies an extraction against a control and against the rendered page, which catches dropped pages, duplicated blocks, and scrambled reading order — but no signal detects a value transposition that preserves the words, the length, and the order. A table extractor that swaps two cells produces this, and every check passes:

control:   Alice status approved   Bob status rejected
candidate: Alice status rejected   Bob status approved

Both claims are inverted; containment, inflation, order, and coverage are all perfect. A table-derived claim needs human verification. Recorded rather than hoped away.

Zero runtime dependencies — Python 3.11+, standard library only.

The two loops

• Loop A — encode expertise (klode.lib): turn sources into cited, retrievable knowledge. Every card claim carries a verbatim (grep: …) anchor; klode check fails if any citation stops resolving.
• Loop B — supervise work (klode.gate): submit a draft, score it against a CriterionSpec — an authored, corpus-pinned, human-approved rubric with stable ids and behaviorally anchored levels — and return Go / Recycle. Each cited defect is grounded through klode.lib.verify, so the judge's citations are un-fakeable, and every authored field declares whether it is the source's words, a paraphrase, an inference, or not stated at all. The boundary is enforced: klode.gate consumes only the klode.lib public API, never its internals. See dev-docs/SPEC-criterion.md.

Install

pip install klode
# or from a source checkout:
pipx install -e .              # provides the `klode` and `klode-mcp` commands

Three surfaces over one engine

A knowledge base is a library.toml plus a corpus of cited cards. Point klode at one with -c, or register several in a manifest and address them by id (--kb <id> / --registry).

CLI — klode:

klode -c path/to/library.toml check                 # citation-rot linter (exit 1 on any broken citation)
klode -c path/to/library.toml search pacing         # retrieval over the cards
klode -c path/to/library.toml consult brevity       # read a craft lens
klode -c path/to/library.toml verify brevity "the exact quote"   # prove a quote against its source
klode ingest paper.pdf --shelf papers               # ingest a source -> clean, grep-ready text
klode kbs                                            # list the KBs in a registry

Add --json to any read verb for machine-readable output (the same structured result the MCP renders).

MCP server — klode-mcp (stdio): exposes the read/verify surface to an agent — list_kbs, search_sources, consult_dimension, consult_framework, zoom_card, verify_quote, diagnose, list_lenses. Serves one KB (-c) or many (--registry), and every grounded result names its KB.

Library — import klode.lib: the stable public-API facade — verify, search, consult, resolve, diagnose, Config. Cheap to import; pulls in no frontends or optional backends.

Ingestion — any format to grep-ready text

klode ingest detects the format by content signature (not the extension) and converges every source on one clean-text pipeline: PDF, EPUB, DOCX, HTML/XHTML, TXT. EPUB/DOCX/HTML/TXT are pure stdlib; only PDF's OCR tiers are optional.

For table-heavy or multi-column PDFs, point klode at a docling-serve endpoint — the GPU runs server-side, so klode stays dependency-free — and pass --tier docling:

# ~/.klode/settings.toml
[ingest]
docling_url = "http://<host>:15001"
tier = "docling"          # or leave at "auto" and escalate only when the text layer is bad

KLODE_DOCLING_URL overrides the file, and klode settings prints every value with the source that won. The URL is topology, not a credential: bind docling-serve to a private interface, and do not rely on the URL being unguessable.

marker is supported the same way (marker_url, --tier marker) and is remote-only — it pulls torch and layout models, which klode does not depend on. It is deliberately not in the auto escalation ladder: a backend earns a ladder slot by measuring better than the one it would displace, not by being installed.

Backends are chosen by measurement, never by intuition — eval/extract_bakeoff.py ranks them against the rendered page (pdftoppm + tesseract), the one signal not downstream of another extractor. The committed result over 20 two-column academic papers (eval/results/extract-bakeoff-2026-08-11.json): docling and pdftotext tie on recall (0.931 vs 0.945), and docling wins decisively on reading order — median 1.000 against 0.697, better on 10 documents of 17 and worse on 1. That is what earns it the tier-3 slot. marker failed 16 of the same 20 documents on the deployment tested, so it has no paired basis to be ranked at all.

Architecture

One operations registry projects a single core to both surfaces (CLI + MCP), so they can differ in formatting but never in behaviour. Every grounded result carries structured provenance (which KB, which source, which policy). See dev-docs/ for the design record and SPEC-operations.md for the machine-readable operation table.

Status

0.4.1 — beta. The engine (klode.lib) is solid: 867 tests, a stable public-API facade, an AST layering guard, a content-sniffing multi-format ingester, and an MCP server, all with zero runtime dependencies. klode.gate (Loop B) is now a fail-closed supervising gate — freshness/review-aware grounding (verify_evidence), a bounded evidence-context op (verify_context), a fail-closed verified-context bundle (build_context_bundle), a shared structured anchor contract (parse_markers/Marker, regex/context/#n), and CriterionSpec v1 as its sole rubric input (field-level epistemics, behaviorally anchored levels, a computed corpus fingerprint, and a human-approval admission gate) are in place, and the rubric judge is real (LLMJudge: G-Eval two-step form-filling, position-bias-debiased by balanced permutation, transport-injectable, stdlib-only). What is not done is calibration: no rubric has yet been measured against human scores, so Verdict.calibrated is False and every verdict is marked non-production — by mechanism, not by convention. The optional entailment check (klode check --entail) pulls a small NLI model behind klode[entail] and is advisory, warn-only.

License

MIT — see LICENSE.