Reasoning degrades when evidence (what is actually observed or verifiable) is blended with inference (what is deduced) and assumption (an unstated premise). Models are especially prone to this: they present fluent inference in the same confident register as fact. This skill separates them: it labels each claim in a body of text as evidence, inference, or assumption, records the basis for each, attaches a confidence level to inferences, and flags anything uncited. The output is an evidence/inference ledger. Note the boundary: this classifies claim type; it does not verify that the evidence is true (that is a separate, fact-checking job).
When to Use
- A recommendation, plan, or conclusion must be trusted before it is acted on.
- High-stakes contexts: legal, medical, financial, safety, architecture and planning.
- Auditing the reasoning behind a conclusion, including the agent's own.
- As a step in a reasoning-audit workflow.
When NOT to Use
- As a fact-checker. It labels what kind of claim something is, not whether it is true.
- On creative or exploratory work where rigor is not the point.
- On trivial claims, where sorting produces only noise.
- When the claims are already well-sourced and the leaps are already explicit.
Instructions
When asked to sort evidence from inference, follow these steps:
- Collect the claims. Break the prompt, document, or proposed conclusion into discrete claims. Keep each to one assertion.
- Label each claim. Mark it Evidence (observed or verifiable, with a source), Inference (deduced from other claims), or Assumption (an unstated premise it depends on).
- Record the basis. For evidence, name the source or observation. For inference, name what it is inferred from. For assumption, state the premise plainly.
- Rate inference confidence. For each inference, assign high / medium / low and say why. Do not treat plausibility as verification.
- Flag the gaps. Mark anything presented as fact but uncited, and any load-bearing assumption that is unexamined.
- Surface the load-bearing unknowns. List the few unsupported claims that most need verification before the conclusion is trusted.
- Emit the ledger per
references/TEMPLATE.md.
Output Format
Use the template in references/TEMPLATE.md. The deliverable is the ledger plus the load-bearing-unknowns list, not prose.
Quality Checklist
Before finalizing, verify:
- Every claim is labeled evidence, inference, or assumption.
- No confident inference is mislabeled as evidence.
- Inferences carry a confidence level with a reason.
- Uncited "facts" and unexamined assumptions are flagged.
- The output does not claim to have verified truth, only sorted claim type.
- The output is the ledger artifact, not prose.
Evidence
Tier P. The evidence/inference distinction is a foundational critical-thinking competence (Facione, Delphi Report 1990), and explicit critical-thinking instruction shows moderate gains broadly (strongest for argument mapping, an adjacent technique). This specific "sort into a ledger" method is practitioner-grade, evidence is transferred from human contexts, and the skill verifies claim type, not truth. Full grading: evidence/dossier.md.
Examples
See references/EXAMPLE.md for a completed ledger.