SandBase Agent Skills v0.3.0 — verifiable multi-source research
Verifiable multi-source research
multi-source-search now ships an offline evidence-ledger validator instead of relying on prose-only confidence guidance.
What changed
- Validates claim, source, provider, confidence, conflict, and evidence-gap structure
- Rejects duplicate URLs, unknown or unused source references, and inflated confidence
- Requires high-confidence claims to cite at least three independent sources
- Treats retrieved pages as untrusted input and documents prompt-injection defenses
- Adds explicit credential, sensitive-data, and read-only boundaries
- Adds a documented JSON report contract and five deterministic tests
- Adds direct skills.sh install links for six flagship workflows
Verification
- 25 public Skill validations passed
- 3 Node tests and 19 Python tests passed
- npm package dry-run verified 196 files
Install the flagship workflow:
npx skills add sandbaseai/sandbase-skills --skill multi-source-search --agent codexNo npm package was published as part of this release.