Skip to content

Enforce cheap-model eligibility from the actual assignment shape #724

Open

Description

@Meta-Develop

Proposal

Supply trusted typed assignment shape to selection and enforce bounded cheap-model capability ceilings against that shape, rather than deriving eligibility from the role alone.

Background and evidence

On main fa0c79b9ae38eca2c59b2b1911becfe1a6b3c9ab, src/supervise/selection_bridge.rs:1738-1765 calls task_profile_for_role. Every Worker is represented as medium risk, bounded, medium context and medium horizon (2534-2550), irrespective of its authored task and paths. Plan entries carry prose/paths without trusted typed shape (src/supervise.rs:1991-2025,2287-2305).

Prior-derived catalog construction (selection_bridge.rs:2374-2437) supplies permissive typed maxima, while limitations of bounded cheap lanes remain prose. The selector checks typed capability fields (src/selection/selector.rs:1067-1083). The pressure test at selection_bridge.rs:6123-6175 demonstrates cheap provisional candidates entering selection from the generic Worker profile.

A broad, high-risk or long-horizon Worker can therefore receive the same selection eligibility input as a localized Worker. This concerns task fitness; existing path, runtime and publication authority gates must remain unchanged.

Acceptance criteria

  • Represent risk, boundedness, context, horizon, authority/effects and delegation in a trusted, validated assignment shape.
  • Use that shape in selection input, replay provenance and outcome-history keys.
  • Encode cheap-lane restrictions as enforced capabilities; unproven shape must use conservative eligibility.
  • Test same-role tight versus broad/high/long assignments, including quota pressure.
  • Operator priors must not silently widen bundled capability ceilings without applicable qualification evidence.

Advantages

Makes low-cost decomposition useful without assuming every task assigned to a Worker is small or equivalent, and keeps learned outcomes comparable.

Disadvantages and costs

Requires schema/provenance migration and a trusted shape derivation/validation boundary. Conservative unknown handling can reduce cheap-lane utilization until evidence is available.

Affected scope and verification

Supervisor plans, selection bridge/catalog construction, selector capability tests, replay and outcome history. Source-level audit only. #242 describes bounded-leaf intent; #631/#688 concern empirical qualification, which is separate from enforcing actual assignment shape. No open Issue/PR covers this input gap.

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

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions