[
  {
    "id": "graph_handoff_solved",
    "category": "transitivity",
    "prompt": "Handoff log: a hands off to b; b hands off to c. Ignore the separate handoff x hands off to b. The handoff chain beginning at a ends at",
    "expected_display_answer": "c",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "graph_pipeline_solved",
    "category": "transitivity",
    "prompt": "Pipeline notes list alpha => beta | beta => gamma | gamma => delta. A decoy note says x => beta. Starting from alpha, report the terminal item:",
    "expected_display_answer": "delta",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "graph_primary_solved",
    "category": "transitivity",
    "prompt": "primary edge n1 -> n2; primary edge n2 -> n3; secondary edge n1 -> z9; secondary edge z9 -> n2. Follow primary edges from n1; answer:",
    "expected_display_answer": "n3",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "graph_audit_solved",
    "category": "transitivity",
    "prompt": "Audit trail says p1 precedes p2, p2 precedes p3. It also says q1 precedes p2, but that trail is unrelated. From p1, the last reachable node is",
    "expected_display_answer": "p3",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "graph_ambiguous_refused",
    "category": "transitivity",
    "prompt": "Graph ledger: main(a,b); main(a,c). Resolve main* from a; terminal node:",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "graph:ambiguous_branch"
  },
  {
    "id": "graph_cycle_refused",
    "category": "transitivity",
    "prompt": "Graph ledger: main(a,b); main(b,a). Resolve main* from a; terminal node:",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "graph:cycle_detected"
  },
  {
    "id": "graph_missing_link_refused",
    "category": "transitivity",
    "prompt": "Pipeline notes list a => b | c => d. Starting from a, report the terminal item:",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "graph:missing_link"
  },
  {
    "id": "graph_blocked_refused",
    "category": "transitivity",
    "prompt": "Dependency map has stable edge a->b; blocked edge b->c. Using stable edges only, a reaches",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "graph:blocked_edge_semantics"
  },
  {
    "id": "arith_ledger_solved",
    "category": "arithmetic",
    "prompt": "Ledger entry: begin 20, pay out 5, add back 3, scale by 2. The ledger closes at",
    "expected_display_answer": "36",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "arith_counter_solved",
    "category": "arithmetic",
    "prompt": "Counter starts at 10. Ops: +5; *3; -4. Counter ends as",
    "expected_display_answer": "41",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "arith_quotient_solved",
    "category": "arithmetic",
    "prompt": "Compute the integer quotient after taking 7 times 4, adding 5, and dividing by 3. Quotient:",
    "expected_display_answer": "11",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "arith_missing_operand_refused",
    "category": "arithmetic",
    "prompt": "Counter starts at 10. Ops: +5; *unknown; -4. Counter ends as",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "arithmetic:missing_operand"
  },
  {
    "id": "arith_divzero_refused",
    "category": "arithmetic",
    "prompt": "Machine value 9 is doubled, then 3 is added, then the result is divided by 0. Integer quotient:",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "arithmetic:division_by_zero"
  },
  {
    "id": "arith_xor_refused",
    "category": "arithmetic",
    "prompt": "Ledger entry: begin 9, apply xor 4, add back 2, scale by 2. The ledger closes at",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "arithmetic:unsupported_operator"
  },
  {
    "id": "code_loop_solved",
    "category": "code_reasoning",
    "prompt": "Trace Python: total=0; for i in range(8): add i only when i is even. total is",
    "expected_display_answer": "12",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "code_augassign_solved",
    "category": "code_reasoning",
    "prompt": "Trace Python: value=4; value += 6; value *= 3; final value is",
    "expected_display_answer": "30",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "code_branch_solved",
    "category": "code_reasoning",
    "prompt": "Trace Python: x=5; y=4; threshold=8; out is 'large' if x+y > threshold else 'small'. out is",
    "expected_display_answer": "large",
    "expected_solved": true,
    "expected_boundary_reason": null
  },
  {
    "id": "code_unknown_function_refused",
    "category": "code_reasoning",
    "prompt": "Trace Python: x=7; y=mystery(x); y is",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "code:unknown_function"
  },
  {
    "id": "code_unbounded_loop_refused",
    "category": "code_reasoning",
    "prompt": "Trace Python: x=7; while x > 0: x = x + 1. x is",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "code:unbounded_loop"
  },
  {
    "id": "code_missing_variable_refused",
    "category": "code_reasoning",
    "prompt": "Trace Python: x=7; y=x+missing; y is",
    "expected_display_answer": "<ABSTAIN>",
    "expected_solved": false,
    "expected_boundary_reason": "code:missing_variable"
  }
]