epistemic-grounding

name	epistemic-grounding
description	Claim-to-evidence grounding discipline for Mycroft skills and recipes. Use when extracting findings, assigning confidence tags, fact-checking claims, ingesting vault notes, or deciding whether a piece of source material is a lead, a partially grounded claim, or a verifiable finding.
version	1.0
invocable_by	["recipe","skill","user"]

Epistemic Grounding

Use this skill whenever Mycroft turns source material into a claim or evaluates whether a claim is ready to file, brief, or publish.

The core question is:

Does this exact evidence justify believing this exact claim?

A source is only an anchor. Grounding is the relationship between the claim and the anchor.

Two Profiles

Mycroft uses this skill at two intensities:

• default — used by morning-brief, vault-qa, obsidian-ingest, source-monitor, content-creator, casual fact-checks. Apply the ladder and confidence-cap rubric in your head; emit only the confidence: high|medium|low tag mandated by SOUL.md plus a source:: line. No required schema.
• fact-check — used by fact-check, fact-check-c2pa, draft-verification work, anything destined for publication or editor review. Emit the full grounding object per the schema in tools/mycroft/docs/grounding-provenance-spec.md and the tools/mycroft/schemas/sift-manifest.schema.json. See the fact-check skill for the contract.

The ladder, confidence caps, and failure router below are universal — they apply at both intensities.

Grounding Ladder

Classify each candidate claim on this ladder:

1. Unsourced signal — interesting, but no source anchor yet.
2. Source-adjacent lead — a source mentions the topic, but does not support the claim.
3. Partially grounded claim — evidence supports some claim elements, but missing assumptions remain.
4. Directly grounded claim — source text directly supports all material claim elements.
5. Independently verified finding — direct grounding plus independent corroboration and no unresolved contradiction.

Only levels 4-5 can be tagged confidence: high. Levels 1-2 are leads, not findings — tag confidence: low, unsourced if you must surface them. Level 3 is at most confidence: medium and often low.

Required Grounding Check

For every claim you intend to write, brief, or publish:

1. Break the claim into material elements: actor, action, object, time, place, amount, relationship, status.
2. Identify the exact quote, table row, record, image frame, metadata field, or document passage that supports each element.
3. Classify the source role:
   • primary — original record, document, direct statement, data source, filing, archived page, or observable artifact.
   • secondary — reporting, analysis, database aggregation, third-party summary.
   • contextual — useful background, but not evidence for the claim.
4. Classify support type:
   • direct — evidence states the claim elements plainly.
   • indirect — evidence supports the claim through a short, explicit inference.
   • inferred — evidence requires unstated assumptions or synthesis across sources.
   • contradicted — reliable evidence conflicts with the claim.
   • insufficient — source does not support the claim.
5. Name missing assumptions and misgrounding risks.
6. Search for contradictions before raising confidence.
7. Apply the confidence cap.

Confidence Caps

Use these caps even if the claim sounds plausible:

Condition	Maximum Confidence
No scraped local file	low
Search snippet only	low
Source is contextual or adjacent	low
Evidence is inaccessible, abstract-only, or materially redacted	low
Claim requires unstated assumptions	medium
Only secondary sources support the claim	medium
Single primary source, no contradiction search yet	medium
Direct primary source plus independent corroboration	high
Credible unresolved contradiction	low or disputed
Output came from a local LLM hallucination-prone for URLs (e.g. fine-tuned Qwen) without fetch verification	low

Never upgrade a claim beyond the weakest material element. If the amount is directly supported but the date is inferred, the whole claim is only partially grounded.

Mycroft-specific notes

• Local-model output is a lead, not authority. Mycroft's default model (fine-tuned Qwen3.5-9B per SOUL.md) hallucinates URLs, citations, statute IDs, and database paths. Any such atom in a model response is confidence: low, unsourced until verified via mycroft-fetch (Crawl4AI/SearXNG) or a qmd lookup. The fine-tune teaches style and domain vocabulary, not citation accuracy.
• unverified ≠ false (from SOUL.md). Evidence-absent is not the same as evidence-contradicts. When grounding is missing, name the gap rather than rewrite the claim into something the evidence supports.
• Apply this skill at ingestion time, not just at fact-check time. A note filed to the vault with a confidence: high tag becomes a citation source for future answers. Inflated confidence at ingest is load-bearing for downstream errors.

When to use the full grounding object

The fact-check profile output schema (7-field grounding block, evidence_refs, human_review state) is required only for the fact-check skill (tools/mycroft/skills/fact-check/SKILL.md) and the fact-check-c2pa.yaml recipe. See tools/mycroft/docs/grounding-provenance-spec.md for the schema and tools/mycroft/schemas/sift-manifest.schema.json for the validator.

default-profile skills do not emit the full object — they apply the ladder and caps in-head and emit the lighter confidence: tag plus a source:: line in vault frontmatter. Promoting routine outputs to the full schema is over-engineering for daily synthesis work.

Failure Routing

When a claim feels wrong, do not patch the wording first. Diagnose the grounding failure:

• Evidence mentions the topic but not the claim: source-adjacent lead.
• Evidence supports a weaker claim: narrow the claim.
• Evidence supports only some elements: mark partial and name missing assumptions.
• Evidence depends on OCR/layout extraction: verify against the original document or image.
• Evidence chains through another citation: trace to the origin.
• Evidence conflicts with another reliable source: mark disputed and preserve both trails.
• Evidence comes only from a fine-tuned local model: re-fetch via mycroft-fetch before raising confidence.

Read references/failure-router.md for deeper failure classes. Read references/grounding-theory.md when designing or revising grounding policy.