skills/advisory-case-director/evals/case_direction_cases.json

{
  "schema_version": "1.0",
  "purpose": "Representative model and human evaluation cases for advisory case direction. Expected semantic outcomes must not be derived by deterministic keywords, a fixed issue tree, or a scorecard.",
  "cases": [
    {
      "id": "initial-market-answer-without-target-data",
      "decision": "Whether to continue diligence on a specialist reseller before target accounts are available",
      "starting_material": "Public evidence suggests unusually cheap procurement of an obsolete commercial product cohort and continuing demand from a narrow user segment. No target purchase ledger, refurbishment data, stock history, or realized resale margin is available.",
      "expected_current_answer": "External evidence can establish whether a plausible and sufficiently durable market profit pool exists. It cannot establish whether the target captures that pool or owns transferable capabilities.",
      "expected_next_branch": "Test the procurement mechanism, technical usability, comparable products and markets, and then derive the few target-data questions needed for a continue or stop decision.",
      "must_not": [
        "Pretend to evaluate target execution",
        "Refuse to form any current answer merely because target accounts are unavailable",
        "Start from a universal due-diligence checklist"
      ]
    },
    {
      "id": "new-evidence-weakens-single-cause",
      "decision": "Whether an unusual contract explains an observed resale spread",
      "prior_answer": "The contract difference is the leading explanation for low acquisition prices.",
      "new_evidence": "Comparable products used a similar contract, but they do not show the same observed resale spread. Concentrated fleet disposals and narrow end-user fit are also documented.",
      "expected_effect": "Weaken the single-cause claim, preserve it as one possible contributor, and revise the causal chain to consider the combined effect of fleet supply, transaction friction, technical usability, and buyer matching.",
      "expected_history": "Keep the prior claim and evidence relationships visible as weakened or superseded rather than deleting them.",
      "must_not": [
        "Keep the answer unchanged because the new source is only external",
        "Replace the complete evidence record with the newest report",
        "Claim that the target captured the spread"
      ]
    },
    {
      "id": "deep-research-answer-opens-target-question",
      "decision": "Whether a historical market profit pool is worth further target diligence",
      "research_return": "The research supports a temporary procurement dislocation and identifies a user segment that can tolerate the product constraint. It finds no evidence that the dislocation persists for newer cohorts and no direct evidence about the target's sourcing or refurbishment economics.",
      "expected_effect": "Condition the answer by product cohort, close the market-mechanism question to the extent supported, and open target questions on purchase lots, all-in acquisition cost, refurbishment yield, time to sale, realized price, and repeatability.",
      "expected_next_branch": "Request the smallest target evidence package capable of testing capture and transferability.",
      "must_not": [
        "Treat historical market evidence as proof of current target margin",
        "Launch broader research without explaining what decision it could change"
      ]
    },
    {
      "id": "partner-question-changes-the-analysis",
      "decision": "Whether the current causal explanation is durable",
      "partner_challenge": "Even if the observed margin is real, why should it persist after the abnormal supply cohort disappears?",
      "expected_effect": "Record the question as partner-originated, identify that it attacks the durability link rather than the existence of historical margin, and redirect the next work toward cohort renewal, replacement sourcing pools, capability transfer, and margin persistence.",
      "must_not": [
        "Attribute the question to Clara",
        "Answer only whether the historical margin was EBITDA or gross margin",
        "Continue the prior work plan without testing durability"
      ]
    },
    {
      "id": "semantic-deck-feedback-round-trip",
      "decision": "Whether to recommend continuing diligence",
      "deck_feedback": "The deck says the market mechanism is proven, but the evidence supports only plausibility and one important alternative explanation remains open.",
      "expected_effect": "Update claim states and the workpaper first, then revise the deck's conclusion, support, uncertainty, and next work.",
      "must_not": [
        "Treat the feedback as wording only",
        "Edit the deck while leaving the spine inconsistent",
        "Use the deck as the case's evidence store"
      ]
    },
    {
      "id": "bounded-data-contribution",
      "decision": "Whether customer concentration makes the current recommendation fragile",
      "branch_request": "Analyze two customer datasets with different periods and definitions to determine concentration, retention, and whether the apparent change is comparable.",
      "expected_handoff": "Give the data specialist the current answer, exact decision question, dataset boundaries, definitions needing review, required calculations and provenance, and the result that would change the recommendation.",
      "expected_return": "Findings, calculation lineage, comparability limits, contradictions, and questions created by the analysis.",
      "must_not": [
        "Make the data orchestrator owner of the overall case thesis",
        "Treat a mechanically valid chart manifest as a business conclusion"
      ]
    },
    {
      "id": "validator-finding-reenters-the-spine",
      "decision": "Whether the completed deliverable is ready to support the continue or stop decision",
      "validation_return": "The validator walks the recommendation upstream and finds that the market mechanism is adequately supported but the target-capture claim is only partially supported. The exact review and audit remain bound to the pre-feedback registers.",
      "expected_effect": "Return the material finding through the validation-feedback envelope, weaken or qualify the active target-capture position, open the smallest target-evidence question, update and checkpoint the workpaper, and rebuild before revalidation.",
      "must_not": [
        "Leave the validator finding in a separate report while the spine remains unchanged",
        "Let deterministic code decide how the finding changes the case answer",
        "Describe the earlier audit as current after the claim or evidence register changes"
      ]
    }
  ]
}