rmems/delegation-planning-trajectories

Delegation Planning Trajectories

Purpose

Long-horizon delegation episodes: a planner splits a task into subtasks, sub-agents execute and report back, and the planner verifies each report before accepting it.

Record schema

JSONL, one record per line. Fields:

record_id, family, task, plan[{subtask_id, description, assignee}], subagent_reports[{subtask_id, claim, evidence}], planner_verification[{subtask_id, check, accepted}], final{outcome, verified}, provenance{produced_at, actors, lineage}

Illustrative example (EXAMPLE ONLY, not real data):

{
  "record_id": "deleg-r01-000001",
  "family": "delegation-planning-trajectory",
  "task": "Add input validation and tests to parse_config()",
  "plan": [
    {
      "subtask_id": "s1",
      "description": "add validation",
      "assignee": "worker-a"
    },
    {
      "subtask_id": "s2",
      "description": "add tests",
      "assignee": "worker-b"
    }
  ],
  "subagent_reports": [
    {
      "subtask_id": "s1",
      "claim": "done",
      "evidence": "diff"
    },
    {
      "subtask_id": "s2",
      "claim": "3 tests pass",
      "evidence": "pytest log"
    }
  ],
  "planner_verification": [
    {
      "subtask_id": "s1",
      "check": "read diff",
      "accepted": true
    },
    {
      "subtask_id": "s2",
      "check": "re-run pytest",
      "accepted": true
    }
  ],
  "final": {
    "outcome": "complete",
    "verified": true
  }
}

Generation

• Open-weight local models only, run via Ollama: gemma4:12b (writer/solver), qwen3-embedding:8b (embedding dedup), FrogNano-4B (Apache-2.0). No hosted frontier models.
• Produced via rmems/synthetic-factory, with provenance gates (every record names its actors, model digests, and seeds) and execution-evidence gates (claims are re-derived from measured execution, not trusted from model transcripts).

Intended use

Post-training of open-weight coding/agent models: supervised fine-tuning (SFT) on trajectories and preference tuning (DPO). Not intended for pre-training or for evaluating hosted models.

Status

7 records (train: 7), generated deterministically by the rmems/synthetic-factory CLI at commit 5e3b048. Every record passed the factory's replay gate; each carries its lane, generate/replay commands, seed and source record id under provenance.lineage. No LLM was used. Field names follow the schema below, with extra fields where the factory emits more detail.

Limitations

• Current records come from deterministic scripted/programmatic generators (no LLM); trajectories may be narrower and simpler than real-world agent behavior.
• Task sources (e.g. public repos) are likely present in base-model pretraining; contamination checks and program-level eval splits are planned but not yet validated.
• English only. Schema may still change in later releases.

Known limitations (this release)

• Small: only 3 tasks in the tool-world-v1 delegation factory.
• Test suites verify the whole tree, so planner_verification[].subtask_id is null; subagent_reports[].evidence is null.
• plan, subagent_reports and planner_verification are extracted mechanically from agent spawn, agent await and run_tests tool calls; the full raw trajectory is kept in trajectory.
• Scripted policy, not a model.
• Tool-call args/arguments are stored as JSON-encoded strings (some episodes deliberately pass wrongly typed arguments, which would otherwise break a fixed column type); json.loads them to recover the original object.