Skip to content

Treat output format as an explicit experimental factor #99

Description

@Colinho22

Motivation

Two diagram representation formats are supported, but format is an implicit choice rather than a controlled variable. Their effects on quality, interpretability, and failure behaviour are unmeasured, so there is no basis for recommending one.

Proposal

  • Make format a first-class factor in run configuration, recorded in telemetry.
  • Score identical tasks under both formats and report the paired delta.
  • Split the failure-cause breakdown by format — a format that is harder to emit validly is a different problem from one that is harder to emit accurately.

Done when:

  • Format selectable per run and present in telemetry
  • Paired comparison reported on matched tasks
  • Failure causes broken down by format

Scope

  • In scope: format as a configuration factor, paired comparison, format-split failure analysis.
  • Out of scope: adding a third format; human-subject evaluation of interpretability.

Open questions

  • Are the two formats expressively equivalent? If one cannot represent a construct the other can, part of the measured quality gap is a representation limit rather than a model limit, and that needs separating.
  • Does each format need its own ground truth, or can one be mechanically converted to the other? Conversion is cheaper but may inject artefacts.
  • Should format interact with strategy in the analysis, or be treated as an additive factor?

Related

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

    enhancementNew feature or request

    Projects

    No projects

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions