Hallucination Regression Hunter

Build eval cases that expose fabricated citations, brittle reasoning chains, and ungrounded tool usage before they hit real workflows.

Source: Advanced first-party pack

Use this skill when

• The request signals

  hallucination

  or a directly related problem.
• The request signals

  grounding

  or a directly related problem.
• The request signals

  eval case

  or a directly related problem.
• The request signals

  agent reliability

  or a directly related problem.
• The likely implementation surface includes

  **/evals/**

  .
• The likely implementation surface includes

  **/tests/**

  .
• The likely implementation surface includes

  **/*.json

  .
• The likely implementation surface includes

  **/*.md

  .

Gather this context first

• Relevant files, modules, or specs that define the current surface.
• Constraints, rollout limits, or non-goals that change the recommendation.
• What success looks like for the user, operator, or release owner.

Recommended workflow

1. Confirm the trigger fit and boundaries before expanding scope.
2. Identify the highest-risk files, interfaces, or decision points first.
3. Produce a bounded plan or implementation slice with exact targets.
4. Run the listed validation hooks or explain what blocks them.
5. Return rollout, fallback, and open-question notes for the next agent or maintainer.

Output contract

• Capability summary and why this skill fits the request.
• Concrete file, module, or artifact targets.
• Validation plan and residual risk notes.

Failure modes to watch

• The pack matches the theme of the request but not the highest-leverage failure domain.
• Validation is mentioned without enough proof for another operator or agent to repeat it.
• The output becomes generic advice instead of a bounded next-step plan.
• Eval or test coverage expands without improving confidence on real failure classes.
• Generated matrices overfit the happy path and miss cross-surface regressions.

Operational notes

• State the smallest safe slice that can be executed or reviewed next.
• Leave enough evidence behind that another maintainer can continue without re-deriving the workflow.
• Call out where human review or approval changes the recommended path.
• Keep the eval loop cheap enough to run during active iteration, not only before release.
• Anchor the test surface in known failure classes before adding breadth.

Dependency and composition notes

• Let this pack lead only when it owns the main bottleneck; otherwise treat it as a specialist sidecar.
• If another pack has a narrower, more concrete surface, hand off with explicit files, risks, and validation goals.
• Pairs well with debugging, routing, and orchestration packs to convert fragile insight into regression coverage.

Validation hooks

• verify_eval_coverage

• verify_text_unchanged

Model chain

• deepseek-ai/deepseek-v3.2

• moonshotai/kimi-k2.5

• deepseek-r1:32b

Pack contents

• SKILL.md

  for portable agent-skill usage

• skill.yaml

  for runtime registry loading

• README.md

  for human install and review context

• marketplace.yaml

  for richer metadata and catalog indexing