Overclaim¶
An ontology of epistemic-overclaim failure modes: named codes for the ways confident language outruns its evidence, in LLM and human prose.
Overclaim is an agent-facing domain. Its primary consumer is a model, not a reader. The site serves a rubric you fetch and apply to a piece of text. It does not judge whether a claim is true; it judges whether the confidence is earned, whether claims are framed honestly, and whether stated certainty matches the evidence shown.
Use it¶
Pick the rubric matching your text's genre, paste it into a context window above the subject text, and apply it:
Fetch https://overclaim.org/rubric/analytical.md
Paste it above the text you want checked.
The model returns: per-claim findings, cleared hedges, and one aggregated decision
(AUTHORIZE < CLARIFY < CONSTRAIN < REVISE < BLOCK).
Or pull the anchor file first and let the agent choose: /llms.txt.
The rubrics¶
Severity is scored relative to the declared frame, so the same line is a defect in an analytical memo and legitimate framing in sales copy.
| Genre | For |
|---|---|
| analytical | memos, reports, design docs, reviews |
| advocacy | proposals, opinion pieces, making a case |
| sales | pitches, landing pages, outbound |
| status | standups, updates, exec replies |
| persuasion | essays, talks, internal selling |
The source of truth¶
The YAML packs are the source of truth; the LLM is the detector; the renderer is a build step that compiles packs into the rubric and never reads subject text. The raw packs are served verbatim for callers that want to compile their own rubric or wire the codes into an eval harness:
/packs/core.yaml, epistemic-overclaim core/packs/substantiation.yaml, the proof-bar axis/packs/grounding.yaml, reasoning-environment warning lights/packs/management.yaml, commitment-without-mechanism
Browse every code in the code catalog, see worked examples of real public text (AI and human) run through the rubric, or read the design for why the ontology is the asset and how the frame gate works.
Status¶
Draft, v1: 39 codes across four packs (core, substantiation, grounding, management). The engineering, sales, and marketing packs are deferred to v1.x until they clear the contribution rule through real use. Ontology content is CC0; code is MIT. The grounding pack is distilled from WillowEmberly / Negentropy (OCAP, CC0).