skills/research-writing/references/research_source_verification.md

# Research and Source Verification

## Research is claim-driven

Start with the claims/questions that must be established, not "find everything about X."

Examples:
- What changed in the current API?
- Which architecture best fits requirement Y?
- What evidence supports benchmark claim Z?
- Is this product still available/priced the same?
- What does the original paper actually report?

## Source hierarchy

Use the strongest source appropriate to the claim.

### Primary / authoritative
- official product documentation;
- standards;
- statute/regulator;
- research paper;
- original repository;
- official pricing;
- source code;
- first-party announcement.

### Strong secondary
- reputable technical journalism;
- independent benchmark;
- respected analyst/practitioner with evidence.

### Community / anecdotal
Useful for real-world friction, undocumented edge cases, sentiment, and practitioner experience.

Community evidence should not silently become a product guarantee.

## Claim ledger

For important research record:

| Claim | Source | Type | Date | Scope/version | Confidence | Caveat |
|---|---|---|---|---|---|---|

## Verification questions

For each material claim:
- Is the entity exact?
- Is the source current?
- Does it describe the same version?
- Is the number measured or marketing?
- Is it first-party or independently reproduced?
- Does the source support the whole sentence?
- What would falsify the claim?

## Time-sensitive research

Especially verify:
- model names;
- APIs;
- SDK behavior;
- pricing;
- product availability;
- benchmarks;
- policies;
- rankings;
- current people/roles.

A source can be reputable and still be stale.

## Conflicts

When evidence conflicts:
1. normalize scope;
2. compare dates;
3. compare definitions;
4. compare versions/populations;
5. prefer authoritative direct evidence when appropriate;
6. preserve unresolved disagreement.

## Negative evidence

"Could not find" is not proof that something does not exist.

Prefer:
- not documented in reviewed sources;
- unable to verify;
- no evidence found in the inspected material.

## Reproduction vs reported result

Keep separate:
- creator/vendor says it worked;
- artifact/source was inspected;
- test was independently reproduced.

## Source register

For large projects record:
- title;
- URL;
- publisher/author;
- publication/update date;
- evidence role;
- supported claims;
- known limitation.

## Stop condition

Research is done when:
- decision-critical claims are supported;
- major contradictions are resolved or disclosed;
- new searches produce diminishing decision value;
- remaining gaps are explicit.