Skillshelf
A controlled task catalog linking seller rights, intended use, input contracts, versions, evaluations, permissions, buyer review, execution evidence, disputes and payment states.
Life-sciences teams may want to procure narrow AI tasks without adopting a large enterprise platform. The supplied research confirms an enterprise scientific-agent product with flexible commercial terms, a free open pharma-agent project and no reviewed self-serve marketplace combining per-task access with evidence packs. The exact marketplace demand, buyer and pricing model remain unverified.
Skillshelf should begin with one low-risk, nonclinical and nonregulatory task rather than five heterogeneous seed verticals. It would preserve seller identity assertion, rights evidence, intended use, prohibited use, input schema, data requirements, model and prompt versions, tool permissions, evaluation set, metric definitions, observed results, limitations, reviewer attestation, buyer acceptance criteria, task request, provider acknowledgment, output, correction, dispute, charge and settlement.
An evaluation report does not validate scientific truth, clinical utility, regulatory compliance or fitness for a buyer's use. Reviewer attestation records who reviewed what; it is not independent assurance. Marketplace sellers may lack rights to data, prompts or methods. Per-task charging can reward volume and discourage uncertainty or escalation. Outputs must remain recommendations or work products for qualified review, never autonomous laboratory, clinical, promotional or regulatory actions.
The pilot should use synthetic or public data and prohibit patient, confidential research and regulated submission data. Marketplace operators need seller due diligence, version freezes, vulnerability response, IP takedown, incident handling, refunds and tenant isolation. The buyer hypothesis is a life-sciences engineering, operations or procurement team buying a bounded task through a marketplace operator; task frequency, integration, budget, risk class and incumbent alternatives need validation.
A life-sciences engineering, operations or procurement team buying one bounded, reviewable AI task through a controlled marketplace operator.
Current enterprise and open-agent activity supports experimentation without a hard procurement window.
A catalog and metered execution scale through software when tasks and reviews are bounded.
The supplied record has one cross-reference and no inbound or broader grounded convergence.
The input identifies a concrete per-task packaging mechanism, confirms enterprise and open alternatives, and reports no specialist self-serve marketplace.
Marketplace liquidity is unverified, the five proposed tasks have different risk regimes, evidence packs can overclaim validation and seller operations add cost.
Discussion
No comments yet — be the first to weigh in.