violetxi/wmrl-v4-note-conditioned-rollouts-20

20 fresh note-conditioned Qwen3.5-9B actor trajectories

These are complete tool-using agent trajectories. Every task executed 2–11 document tools: 56 read, 11 grep, and 2 glob calls in total. No generated reasoning, executed tool call, returned observation or final answer was removed. The base actor system is byte-identical to the published base agentic eval system; the teacher-only memory instruction and notes were appended to it.

The default table begins with tool_sequence, transcript, and messages. transcript shows the task and all turns chronologically. messages contains the exact teacher conversation, including its system prompt and actual tool results (represented as user-role <tool_result> messages by the eval protocol). actor_system exposes the unchanged eval prompt; teacher_system includes the added memory block. events_json preserves the complete structured event log. The final_answer and final_reasoning columns are convenient extracts, not the complete trajectory. Readable example pages show tool calls and the full trajectory first, with the supplied notes at the end.

Steps 1–5 of the note-conditioned procedure: original tasks plus their associated frozen notes, followed by fresh actor execution on the pinned DMS. This is teacher data for qualitative inspection. No student training export, semantic filtering, correctness grading or post-hoc reasoning rewrite has been performed.

• —Trajectory table: 20 rows, with task, supplied notes, full transcript, final reasoning and answer.
• —Turn table: 101 rows, with reasoning, action and returned observation side by side.
• —Note table: 45 task–note rows, with exact note text and read/write provenance.

How these were generated

The same 20 original tasks used in the earlier rewrite comparison were recovered without replacement from violetxi/wmrl-v4-note-sessions at b2b72c696a88e45874a3886b3941c0897f0661a6. All 45 associated notes resolved in violetxi/wm-rl-notes-v4 at 6e62edba847d8013a5e7aefb1e7fd19974fa5381. Associations require actual successful reads or accepted writes; source completion status was not used to infer note reading or useful note use.

Each teacher receives the actor system, original task, and its complete note bundle as prior remembered knowledge. Notes are included verbatim. The teacher can call only glob, grep and read; no note tools or note-writing gates are available. It chooses fresh actions, including answering directly when memory suffices. No original reasoning, action sequence, observations or final answer is added to its generation input.

Teacher: Qwen/Qwen3.5-9B, revision c202236235762e1c871ad0ccb60c8ee5ba337b9a; thinking enabled; temperature 0.6, top-p 0.95, top-k 20, presence penalty 0; up to 20 assistant turns, 16,384 output tokens per turn and 65,536 total context tokens. The native Qwen template is used. Prior assistant reasoning is preserved in the raw records and tables, although the native template omits it from later model prompts after user-role tool results. The note memory remains available throughout.

Hardware recovery: 17 trajectories completed entirely on the local pinned model. GPU 2 failed with three requests in flight and prevented CUDA initialization even when excluded. The three unfinished tasks were resumed from byte-identical saved response history using Qwen3.5-9B hosted by Together through Hugging Face. No completed local model output was replaced. The provider reports weight_version=default and does not expose an immutable weight commit; the exact hosted revision/precision is unverified. providers, model_revision and per-turn provider distinguish these continuations. They are not a strictly identical-backend comparison. Decoding values above are the requested settings; hosted prompt token counts and raw provider metadata are retained. For recovered rows, wall_seconds measures the successful resumed run, excluding outage time.

All tools execute on the reconstructed 9,288-document DMS at Harvey Labs commit c2488cfa24fd01ee88016a121479a2f86b394bd4. All documents parsed successfully; 53 original reads matched the reconstruction byte for byte. New tool observations were independently reexecuted for export verification. Complete assistant text was checked against raw model responses. Restoring a server-stripped </tool> stop delimiter is the only assistant-text serialization addition.

Reading the data

Status counts: {"done": 20}. There are 69 executed tool calls and 60,425 generated tokens. A done status means a closed final answer was emitted within the budget; it does not establish task correctness or beneficial note use. There were 0 tool-error observations and 12 turns with harness feedback, retained in the transcript and turn table. Partial or defective outputs, if present, are included with their actual status.

Notes written while solving the original task are intentionally available at the start of its fresh replay, and may contain its answer. This is knowledge rehearsal, not a causal reconstruction of the original session or held-out evaluation. The sample was selected for the earlier rewrite investigation; it is not random or representative. Check factual fidelity, how memory informs decisions, and whether the final reasoning supports the actual answer during qualitative review.

memory_notes contains privileged teacher input. transcript contains the original task and the complete fresh reasoning/actions/results. final_reasoning and final_answer come from the same generated final turn. All text is model-authored and unedited. Native responses and exact teacher messages are retained locally; the complete teacher messages and execution events are also in raw/<task_id>.json.

Provenance and files

Manifest, structural validation, actor system, teacher memory instruction. The underlying C&H environment is synthetic, from Harvey Labs (MIT); the source note/session datasets are MIT. These records contain synthetic firm tasks and documents, not real client material. Model licensing remains governed by Qwen's model license. Code snapshots and JSONL versions accompany the tables.