Skip to content

[M8][contributor] Add synthetic placeholder/protocol eval fixtures and deterministic checks #13

Description

@AllenXiao95

Parent roadmap

Contributes to #9 ([M8] Add reproducible i18n Prompt evaluation harness and baseline reports).

Why this is a good contributor task

This task is intentionally narrow: it does not require implementing model judging, Prompt authoring, AgentRun orchestration, or release integration. The goal is to establish a small, legally reusable deterministic regression slice for the future i18n evaluation harness.

Scope

Add a small synthetic evaluation fixture set covering response protocol and placeholder preservation. Cases should be original synthetic examples and include provenance/license metadata compatible with public redistribution.

Cover at least:

  • printf placeholders such as %s, %d, positional variants where supported
  • Python-style placeholders such as {name} / {count}
  • escaped percent/braces where relevant
  • multiple placeholders in one source string
  • reordered placeholders when the format permits it
  • missing, duplicated, or mutated placeholders as negative outputs
  • protocol completeness examples such as missing items, extra items, malformed terminators, or structurally incomplete output

Implement deterministic checks that can report structured pass/fail evidence for these cases. Keep the implementation independent of any specific LLM provider.

Suggested artifact shape

The exact schema may evolve with #9, but fixtures should preserve the fields needed by the M8 design, including:

  • stable case ID
  • source/target locale
  • source text and optional context
  • required tokens/assertions
  • expected deterministic outcome
  • provenance kind and license
  • tags/difficulty

Non-goals

  • No semantic or style model judge.
  • No network/model API calls.
  • No Prompt baseline comparison.
  • No game text copied from sources with uncertain redistribution rights.
  • No release gate integration.

Acceptance criteria

  • Fixtures are synthetic/publicly redistributable and include provenance/license metadata.
  • At least 12 focused cases cover valid and invalid placeholder/protocol behavior.
  • Deterministic checks identify missing/extra/mutated placeholders and protocol-structure failures.
  • Failed checks return structured evidence identifying the case and failed assertion.
  • Tests require no provider credentials or network access.
  • Tests pass on supported Windows and Linux CI environments.
  • Contributor documentation explains how to add one additional case without changing evaluator code.

Coordination

If #9 lands a finalized manifest/case schema first, use it directly. If this task lands first, keep fixture structures small and easy to migrate rather than building a competing evaluation framework.

Activity

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

Metadata

Metadata

Assignees

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions