Proposal: bounded multi-source research skill with an auditable evidence ledger

Proposal

Would the maintainers be interested in a multi-source-research Skill that teaches coding agents to investigate a claim across independent sources, stop under an explicit retrieval budget, record conflicts, and validate a machine-readable evidence ledger before synthesis?

The gap it fills is different from the existing development, document, media, and vision Skills: it is a reusable research-control workflow rather than another content generator or API wrapper.

Intended behavior

1. Define the question, evidence standard, freshness window, and search budget.
2. Prefer primary sources and require source independence before raising confidence.
3. Record each source, claim, support/contradiction relation, and retrieval status.
4. Stop when the budget is exhausted or evidence thresholds are met; never hide unresolved uncertainty.
5. Run an offline validator that rejects unknown/duplicate sources, unused evidence, inflated confidence, and high-confidence conflicted claims.
6. Produce both a readable answer and the auditable ledger.

The Skill would remain vendor-neutral and use search/page/browser tools exposed by the host agent. It would not require a SandBase account or MiniMax API beyond whichever model/tool environment the user already runs.

Existing reference implementation

SandBase Skills already ships an Apache-2.0 implementation with a worked example and validator:

• Skill
• Worked example
• Example evidence ledger

If this scope is welcome, I can adapt it to this repository’s skills/<name>/ structure, MIT licensing expectations, concise context budget, README sync, and validation rules as a single-purpose contribution. Attribution would be retained as required by the source license.

Disclosure: I help maintain SandBase Skills. I am asking about scope before preparing an external-repository code contribution, and I am not claiming MiniMax endorsement.