Skip to content

[Layer 5] Who defines the reference task behind acmp:price_basis? #7

Description

@marc-poljak

Context. Layer 5 § Open Questions: "Who defines the reference task behind acmp:price_basis, and who may publish or version such a catalogue?"

Layer 5 advertises an indicative price per capability (acmp:price_cu_indicative), measured on a reference task named by acmp:price_basis. Without a shared definition of that task, two providers' figures describe different work — and the field is no more comparable than the per-token price it replaced.

Your task. Work out who could own such a catalogue, and what it would take:

  • Who publishes it? The working group, an independent body, competing catalogues?
  • What goes in it? What makes a task a fair reference for a capability like sentiment-analysis — input size, difficulty spread, scoring?
  • Versioning. The example value is acmp:reference-tasks/v1. What happens to advertised prices when v2 lands?
  • Prior art. How do MLPerf, SPEC, or benchmark suites in other markets govern their definitions?

This is the discovery-side face of the CU tier-definition question (RFC-0001 §7, question 2) — its candidate Assessor role maintains benchmark definitions, and a reference task is one. An answer here may constrain that one.

Done = a comment or PR weighing at least two governance models against ACMP's principle of provider-agnostic simplicity (P3).

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    good first issueGood for newcomersquestionFurther information is requestedwg-discoveryLayer 5 — Discovery (ARD binding) working group

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions