skills/horizon-forecast/references/evidence-method.md

# Evidence and claim audit

## Source record

For each source keep: source ID, original URL/path, title, publisher/author, publication date, retrieval date, observation period, original data lineage, geography, market definition, units, currency basis, incentives, access status, and a short relevant finding. Include page/table/section where possible. Keep excerpts short; prefer accurate paraphrase.

For each material claim keep: claim ID, exact statement, type (observation/calculation/assumption/forecast), source IDs, support relationship (supports/contradicts/context only), uncertainty, freshness, importance, and verification status. Mark whether you opened and inspected the underlying content. Search snippets and blocked pages cannot certify a claim.

## Evidence strength

Assess directness, methodology, transparency, independent corroboration, timeliness, representativeness, and incentives. Primary does not mean unbiased: an audited total may be strong evidence for past revenue while management guidance remains a forecast. A peer-reviewed proof of concept does not establish commercial throughput or willingness to pay.

Useful confidence labels:

- **High:** pivotal claims have direct and current-enough evidence; definitions align; major mechanisms are observable; plausible alternatives have been checked. High confidence in inputs is not certainty about a future event.
- **Medium:** several credible observations support the thesis, but material assumptions, a short track record, or a transfer from other regions remains.
- **Low:** indirect signals, concentrated source lineage, uncertain definitions, unproven adoption, or unresolved material conflicts dominate.
- **Insufficient:** one or more pivotal claims cannot be established; withhold a definitive ranking or size estimate.

Explain the label in a sentence tied to actual evidence, not a mechanical source count. A likely event may have low confidence; an unlikely event may have high confidence.

## Contradiction protocol

Before averaging incompatible numbers, compare boundaries, measurement dates, currencies, nominal/real basis, annual/cumulative basis, sold/installed/announced state, and primary lineage. Reconstruct comparable definitions if possible. If a difference remains, report the alternatives and test whether it changes the result. Do not resolve disagreement by selecting the largest estimate or the most prestigious-looking publisher.

## Search patterns

Use decision-specific queries: buyer + budget + adoption/renewal; product + realized price + unit cost; supply + commissioned capacity + utilization; field + cancellation/failure + reason; regulator + official decision + jurisdiction. Trace forecasts back to inputs. Search non-English official sources when material, checking translation uncertainty for key terms.

An underappreciation thesis needs a comparator: coverage relative to adjacent fields; investment relative to required capacity; deployment relative to documented demand; or a demonstrably neglected bottleneck. Sparse search results alone do not establish an opportunity. A heavily funded field can still have overlooked niches; a neglected field can be uneconomic for good reasons.

## Audit priority

Verify every claim whose reversal changes a top ranking, event probability materially, or a conclusion about market size. Sample lower-impact supporting claims. When evidence is weak, narrow the claim, widen the range, reduce confidence, or abstain. Do not create fake citations, institutions, precision, or expert endorsements.