Distributed MCTS Trace Event Dictionary

Field-level contract for distributed MCTS trace lines (*_mcts_trace.jsonl).

Scope and Enablement

This dictionary applies when both are true:

--trace-mcts
--mcts-mode distributed

Typical files:

<theorem>/wild_type_mcts_trace.jsonl
<theorem>/<intervention>_mcts_trace.jsonl

Mental Model

Distributed runner agents execute concurrently, but trace output is serialized by global iteration:

                +----------------------------+
                | global iteration allocator |
                +-------------+--------------+
                              |
      +-----------------------+-----------------------+
      |                       |                       |
   agent 0                 agent 1                 agent 2
      |                       |                       |
      +---------- reserve / expand / backprop -------+
                              |
                    OrderedIterationWriter
                              |
                    *_mcts_trace.jsonl (JSONL)

Reservation and intervention flow per agent:

select candidate node
      |
      v
blocked or delayed?
  |                |
 yes              no
  |                |
record reason=     reserve inflight
blocked|delayed    then expand
                   |
                   v
        reason=expanded|no_expansion|no_goal|dead_node|terminal_node

Ordering and Stability Guarantees

iteration is a global counter across all agents, not per-agent.
Output lines are written in ascending iteration order.
Agent-local scheduling step (schedule_iteration) is not logged.

Event Kinds

Two top-level event values are emitted:

iteration: primary per-iteration record
abort: out-of-band abort marker emitted when progress callback aborts search

The distributed runner can emit both in the same global iteration:

{"event":"abort", ...} (out-of-band marker)
{"event":"iteration","reason":"abort", ...} (agent record)

`event = "iteration"` Top-Level Fields

Field	Type	Required	Meaning
`event`	string	yes	Always `"iteration"`
`iteration`	int	yes	Global iteration id
`reason`	string	yes	Iteration outcome category
`agent_id`	int	yes (current impl)	Agent that produced the record
`selected_path`	`list[str]`	yes	Root-to-selected node mvar path
`node`	object	yes	Snapshot of selected node state
`tactics`	`list[object]`	yes	Candidate tactics in attempted order
`attempts`	`list[object]`	yes	Attempt results captured for this iteration
`expanded`	bool or null	yes	Whether any tactic expanded/committed
`terminal_reached`	bool	yes	Whether terminal state was reached
`backprop_success`	bool	yes	Whether this iteration backpropagated success
`tree`	object	yes	Tree-level snapshot after iteration handling
`block`	object	optional	Present on blocked records
`delay`	object	optional	Present on delayed records
`reroute`	`list[object]`	optional	Reroute trail before final node/decision

Note: trace_context fields (currently tier, budget) are merged into each record by caller code.

`reason` Values

`reason`	When emitted
`expanded`	At least one tactic was committed and tree expanded (possibly terminal)
`no_expansion`	Candidate tactics existed but none produced novel expansion
`blocked`	Node selection was blocked by block schedule and not rerouted
`delayed`	Node selection was delayed by delay schedule and not rerouted
`terminal_node`	Selected node already terminal at reservation time
`dead_node`	Selected node already dead at reservation time
`no_goal`	Adapter returned no goal for selected node
`abort`	Progress callback requested abort (agent iteration record)

Nested Objects

`node`

{
  "mvar_id": "cp1:_uniq.560",
  "goal_type": "P -> Q",
  "goal_sig": "sig_abc123",
  "goal_sig_strict": "strict_def456",
  "visit_count": 12,
  "success_count": 5,
  "is_terminal": false,
  "is_dead": false,
  "depth": 3
}

`tree`

{
  "nodes": 41,
  "expansions": 19,
  "max_depth": 6,
  "solved": false,
  "aborted": false,
  "inflight": 2
}

`attempts[]`

Serialized from TacticAttempt:

{
  "tactic": "intro h",
  "outcome": "success",
  "child_mvar_ids": ["cp1:_uniq.561"],
  "timestamp_ms": 1730,
  "tactic_norm": "intro h",
  "goal_sig": "sig_abc123",
  "goal_sig_strict": "strict_def456",
  "goal_type": "P -> Q",
  "peg_id": null,
  "peg_kind": null,
  "block_reason": null,
  "provider_id": "deepseek"
}

outcome is one of:

success
failure
blocked

`block`

{
  "until": 140,
  "remaining": 6,
  "duration": 20,
  "immovable": false
}

`delay`

{
  "until": 92,
  "remaining": 3,
  "duration": 5
}

`reroute[]`

Each entry is a skip decision before final handling:

{
  "mvar_id": "cp1:_uniq.574",
  "reason": "blocked",
  "block": {"until": 140, "remaining": 6, "duration": 20, "immovable": false}
}

or:

{
  "mvar_id": "cp1:_uniq.601",
  "reason": "delayed",
  "delay": {"until": 92, "remaining": 3, "duration": 5}
}

`event = "abort"` Shape

Out-of-band abort marker:

{
  "event": "abort",
  "iteration": 312,
  "tier": 2,
  "budget": 200,
  "tree": {
    "nodes": 141,
    "expansions": 77,
    "max_depth": 11,
    "solved": false,
    "aborted": true,
    "inflight": 4
  }
}

This event does not include agent_id, reason, attempts, or node.

Worked Record Examples

Expanded Iteration with Prior Reroutes

{
  "event": "iteration",
  "iteration": 85,
  "reason": "expanded",
  "agent_id": 2,
  "tier": 1,
  "budget": 50,
  "selected_path": ["cp1:_uniq.560", "cp1:_uniq.568"],
  "node": {
    "mvar_id": "cp1:_uniq.568",
    "goal_type": "A -> B",
    "goal_sig": "sig_f08f",
    "goal_sig_strict": "strict_817f",
    "visit_count": 4,
    "success_count": 2,
    "is_terminal": false,
    "is_dead": false,
    "depth": 1
  },
  "tactics": [
    {"tactic": "intro h", "score": 0.61},
    {"tactic": "simp", "score": 0.23}
  ],
  "attempts": [
    {
      "tactic": "intro h",
      "outcome": "success",
      "child_mvar_ids": ["cp1:_uniq.602"],
      "timestamp_ms": 519,
      "tactic_norm": "intro h",
      "goal_sig": "sig_f08f",
      "goal_sig_strict": "strict_817f",
      "goal_type": "A -> B",
      "peg_id": null,
      "peg_kind": null,
      "block_reason": null,
      "provider_id": "reprover"
    }
  ],
  "expanded": true,
  "terminal_reached": false,
  "backprop_success": true,
  "tree": {
    "nodes": 33,
    "expansions": 14,
    "max_depth": 6,
    "solved": false,
    "aborted": false,
    "inflight": 3
  },
  "reroute": [
    {
      "mvar_id": "cp1:_uniq.574",
      "reason": "blocked",
      "block": {"until": 93, "remaining": 8, "duration": 20, "immovable": false}
    }
  ]
}

Blocked Iteration (No Reservation)

{
  "event": "iteration",
  "iteration": 86,
  "reason": "blocked",
  "agent_id": 1,
  "selected_path": ["cp1:_uniq.560", "cp1:_uniq.574"],
  "node": {
    "mvar_id": "cp1:_uniq.574",
    "goal_type": "A -> B",
    "goal_sig": "sig_f08f",
    "goal_sig_strict": "strict_817f",
    "visit_count": 3,
    "success_count": 1,
    "is_terminal": false,
    "is_dead": false,
    "depth": 1
  },
  "tactics": [],
  "attempts": [],
  "expanded": null,
  "terminal_reached": false,
  "backprop_success": false,
  "tree": {
    "nodes": 33,
    "expansions": 14,
    "max_depth": 6,
    "solved": false,
    "aborted": false,
    "inflight": 2
  },
  "block": {
    "until": 93,
    "remaining": 7,
    "duration": 20,
    "immovable": false
  }
}

Quick Query Patterns

Top reason frequencies:

jq -r 'select(.event=="iteration") | .reason' theorem/wild_type_mcts_trace.jsonl | sort | uniq -c

Average inflight by reason:

jq -s '
  map(select(.event=="iteration" and .tree.inflight != null))
  | group_by(.reason)
  | map({reason: .[0].reason, avg_inflight: ((map(.tree.inflight)|add)/length)})
' theorem/wild_type_mcts_trace.jsonl

Distributed MCTS Trace Event Dictionary

Scope and Enablement

Mental Model

Ordering and Stability Guarantees

Event Kinds

event = "iteration" Top-Level Fields

reason Values

Nested Objects

node

tree

attempts[]

block

delay

reroute[]

event = "abort" Shape