How CONFIG.md evolved across versions.
Added an adversarial verification pass that re-checks each cited claim against a second independent source before it ships.
# Research Rabbit — Config (v2)
## Pipeline
1. Search the web for the question
2. Read the top results, draft a cited brief
3. Verify each cited claim against a second independent source
4. Drop or flag claims that fail verification
## Model
- Claude Sonnet (verification pass on Claude Haiku)
## Notes
- Second-source verification caught confidently-wrong single-source claims
Single pass: search the web, read the top results, write a brief.
# Research Rabbit — Config (v1)
## Pipeline
1. Search the web for the question
2. Read the top results
3. Write a cited brief
## Model
- Claude Sonnet
## Known issues
- Cites the source it read even when the source itself is wrong or dated
Diff CONFIG.md between any two captured versions — not just the latest change.
- # Research Rabbit — Config (v1)+ # Research Rabbit — Config (v2) ## Pipeline 1. Search the web for the question- 2. Read the top results- 3. Write a cited brief+ 2. Read the top results, draft a cited brief+ 3. Verify each cited claim against a second independent source+ 4. Drop or flag claims that fail verification ## Model- - Claude Sonnet+ - Claude Sonnet (verification pass on Claude Haiku) - ## Known issues- - Cites the source it read even when the source itself is wrong or dated+ ## Notes+ - Second-source verification caught confidently-wrong single-source claims