b09501048/schema-robo-c4-opus

Schema-Robo c4: Claude Opus 5.5 runs on Adroit (Pen, Door, Hammer, Relocate)

Every Claude Opus 5.5 campaign of the Schema-Robo c4 experiments: an LLM agent writes a policy.py controller for an Adroit hand task, learns from its own rollouts, submits frozen policies for evaluation on training positions, and the final policy is scored on held-out starts/targets it never saw. 35 campaigns, one archive each (25 GB compressed). Code: GitHub ZhiqiChen05/Schema-Robo-v6-c4 (branches c4-ablations, c4-grasp-general).

Files

Path	What it is
archives/<campaign>.tar.gz	One campaign; extracts to experiment_results/<campaign>/
index.csv	One row per (campaign, task): archive, setup, method, status, held-out successes
manifest.json	Per archive: size, sha256, file count, source worktree
pack_opus_results.py, make_index.py	The scripts that produced the archives and the index

Setups and methods

Setup (campaign prefix)	Stages	Held-out
opus55_po_seq_*	3 seeds learned in sequence	10 episodes
opus55_po_seq10_*	10 seeds learned in sequence, held-out after every stage	30 episodes
opus55_po_pool10_*	1 stage holding all 10 training positions	30 episodes

Method	Mode suffix	What changes
A0	(none) / policy_only	built-in optimizer (optimize_policy)
A12	A12	submit-first protocol
A12+A9	A12A9	submit-first + search budget
L3	L3	no built-in optimizer; model-driven search (evaluate_candidates, sandboxed run_compute)
L3+papers	L3P	L3 + task reference papers in papers/
+ grasp prompt (Pen-specific)	L3PG (*_L3PG_pen_r1)	L3+papers + a Pen-specific grasp-aware controller paragraph
+ grasp prompt (task-general)	L3PG (*_L3PG_general_*)	L3+papers + a task-independent, human-like grasp-aware paragraph
+ critic	L3PGC	Pen-specific grasp prompt + an Opus critic gating submit_to_eval (_v2_: structure-only rubric; the run without _v2_ is the stopped v1 pilot)

Results

Held-out successes per campaign and task (from index.csv):

Setup	Task	Method	Campaign	Held-out
pool10	Door	L3+papers + task-general grasp prompt	opus55_po_pool10_L3PG_general_door_r1	30/30
pool10	Door	L3+papers (model-driven search + reference papers)	opus55_po_pool10_L3P_r1	26/30
pool10	Door	A0 (built-in optimizer)	opus55_po_pool10_r1	30/30
pool10	Hammer	L3+papers + task-general grasp prompt	opus55_po_pool10_L3PG_general_hammer_r1	28/30
pool10	Hammer	L3+papers (model-driven search + reference papers)	opus55_po_pool10_L3P_r1	30/30
pool10	Hammer	A0 (built-in optimizer)	opus55_po_pool10_r1	30/30
pool10	Pen	L3+papers + Pen-specific grasp prompt + critic (v1 rubric pilot, stopped)	opus55_po_pool10_L3PGC_pen_r1	— (stopped)
pool10	Pen	L3+papers + Pen-specific grasp prompt + critic (v2 rubric)	opus55_po_pool10_L3PGC_v2_pen_r1	12/30
pool10	Pen	L3+papers + Pen-specific grasp prompt + critic (v2 rubric)	opus55_po_pool10_L3PGC_v2_pen_r2	7/30
pool10	Pen	L3+papers + task-general grasp prompt	opus55_po_pool10_L3PG_general_pen_r1	17/30
pool10	Pen	L3+papers + Pen-specific grasp prompt	opus55_po_pool10_L3PG_pen_r1	12/30
pool10	Pen	L3+papers (model-driven search + reference papers)	opus55_po_pool10_L3P_r1	6/30
pool10	Pen	A0 (built-in optimizer)	opus55_po_pool10_r1	10/30
pool10	Relocate	L3+papers + task-general grasp prompt	opus55_po_pool10_L3PG_general_relocate_r1	20/30
pool10	Relocate	A0 (built-in optimizer)	opus55_po_pool10_relocate_A0_r1	30/30
pool10	Relocate	L3+papers (model-driven search + reference papers)	opus55_po_pool10_relocate_L3P_r1	24/30
seq10	Door	A12+A9 (submit first + search budget)	opus55_po_seq10_A12A9_r1	28/30
seq10	Door	A12 (submit first)	opus55_po_seq10_A12_r1	30/30
seq10	Door	L3+papers (model-driven search + reference papers)	opus55_po_seq10_L3P_r1	30/30
seq10	Door	A0 (built-in optimizer)	opus55_po_seq10_r1	12/30
seq10	Hammer	A12+A9 (submit first + search budget)	opus55_po_seq10_A12A9_r1	30/30
seq10	Hammer	A12 (submit first)	opus55_po_seq10_A12_r1	30/30
seq10	Hammer	L3+papers (model-driven search + reference papers)	opus55_po_seq10_L3P_r1	30/30
seq10	Hammer	A0 (built-in optimizer)	opus55_po_seq10_r1	30/30
seq10	Pen	A12+A9 (submit first + search budget)	opus55_po_seq10_A12A9_r1	12/30
seq10	Pen	A12 (submit first)	opus55_po_seq10_A12_r1	10/30
seq10	Pen	L3+papers (model-driven search + reference papers)	opus55_po_seq10_L3P_r1	22/30
seq10	Pen	A0 (built-in optimizer)	opus55_po_seq10_r1	12/30
seq10	Relocate	A0 (built-in optimizer)	opus55_po_seq10_relocate_A0_r1	28/30
seq10	Relocate	A12+A9 (submit first + search budget)	opus55_po_seq10_relocate_A12A9_r1	27/30
seq10	Relocate	A12 (submit first)	opus55_po_seq10_relocate_A12_r1	30/30
seq10	Relocate	L3+papers (model-driven search + reference papers)	opus55_po_seq10_relocate_L3P_r1	30/30
seq	Door	A0 (built-in optimizer)	opus55_po_seq_r1	9/10
seq	Door	A0 (built-in optimizer)	opus55_po_seq_r2	4/10
seq	Door	A0 (built-in optimizer)	opus55_po_seq_r3	10/10
seq	Hammer	A0 (built-in optimizer)	opus55_po_seq_r1	9/10
seq	Hammer	A0 (built-in optimizer)	opus55_po_seq_r2	10/10
seq	Hammer	A0 (built-in optimizer)	opus55_po_seq_r3	9/10
seq	Pen	A12+A9 (submit first + search budget)	opus55_po_seq_A12A9_pen_r1	3/10
seq	Pen	A12+A9 (submit first + search budget)	opus55_po_seq_A12A9_pen_r2	3/10
seq	Pen	A12+A9 (submit first + search budget)	opus55_po_seq_A12A9_pen_r3	3/10
seq	Pen	A12 (submit first)	opus55_po_seq_A12_pen_r1	2/10
seq	Pen	A12 (submit first)	opus55_po_seq_A12_pen_r2	1/10
seq	Pen	A12 (submit first)	opus55_po_seq_A12_pen_r3	3/10
seq	Pen	L3+papers (model-driven search + reference papers)	opus55_po_seq_L3P_pen_r1	1/10
seq	Pen	L3+papers (model-driven search + reference papers)	opus55_po_seq_L3P_pen_r2	1/10
seq	Pen	L3+papers (model-driven search + reference papers)	opus55_po_seq_L3P_pen_r3	4/10
seq	Pen	L3 (model-driven search)	opus55_po_seq_L3_pen_r1	4/10
seq	Pen	L3 (model-driven search)	opus55_po_seq_L3_pen_r2	1/10
seq	Pen	L3 (model-driven search)	opus55_po_seq_L3_pen_r3	2/10
seq	Pen	A0 (built-in optimizer)	opus55_po_seq_r1	3/10
seq	Pen	A0 (built-in optimizer)	opus55_po_seq_r2	0/10
seq	Pen	A0 (built-in optimizer)	opus55_po_seq_r3	4/10

Download

pip install -U huggingface_hub
# everything (~25 GB)
hf download b09501048/schema-robo-c4-opus --repo-type dataset --local-dir schema-robo-c4-opus
# or only what you need, e.g. the single-stage runs
hf download b09501048/schema-robo-c4-opus --repo-type dataset --local-dir schema-robo-c4-opus \
  --include "index.csv" "manifest.json" "archives/opus55_po_pool10_*"

Check an archive against manifest.json:

cd schema-robo-c4-opus
python - <<'PY'
import hashlib, json
for name, e in json.load(open("manifest.json")).items():
    try:
        h = hashlib.sha256(open(e["archive"], "rb").read()).hexdigest()
    except FileNotFoundError:
        continue
    print("OK " if h == e["sha256"] else "BAD", name)
PY

Extract

tar -xzf archives/opus55_po_pool10_L3PG_general_pen_r1.tar.gz        # -> experiment_results/<campaign>/
for f in archives/*.tar.gz; do tar -xzf "$f"; done                   # all of them

Only part of an archive (no need to unpack trajectories):

A=archives/opus55_po_pool10_r1.tar.gz
# results only: campaign status, held-out summaries and per-episode outcomes
tar -xzf $A --wildcards '*/campaign.json' '*/report.json' '*/heldout/*/evaluation.json' '*/heldout/*/episodes/*/result.json'
# the final (held-out) policy of each task
tar -xzf $A --wildcards '*/heldout/*/frozen_source/policy.py'
# what the agent was told, and its learning-time accounting
tar -xzf $A --wildcards '*/attempt_01/prompt.txt' '*/attempt_01/fewshot_state.json'

Layout of one campaign

experiment_results/<campaign>/
  campaign.json                 status, tasks, modes, model, job list
  report.json                   campaign summary written at the end
  inputs/<job>.json             resolved config the run used; <job>_heldout.json = held-out case list
  logs/<job>.log                coordinator log
  source_snapshot/              exact source code the run imported
  runs/<job>/runs/repetition_001/case-NN/attempt_01/   one learning stage (seq: case-01..03, seq10: case-01..10, pool10: case-01)
    prompt.txt                  the full prompt given to the agent
    fewshot_state.json          budgets used (provider_requests, tool_calls, search_rollouts, resets) and every submission
    submissions/submission_NNNN/  frozen policy.py + per-position evaluation of each submit_to_eval
    learning/episode_NNNN/      learning episodes (actions, observations) and the CLI session transcript
    search/search_NNNN/         each evaluate_candidates / optimize_policy batch
    scratch/                    the agent's own notes and scripts (L3 modes)
    session_live/               Claude CLI session transcript (jsonl)
    events.jsonl                tool-call event log
  heldout/<job>/                held-out evaluation of the final policy
    evaluation.json             summary: successful_episodes / planned_episodes
    frozen_source/policy.py     the policy that was evaluated
    episodes/heldout-NN/        result.json (success, seed, diagnostics) and trajectory.jsonl (every step)
    videos/                     rendered GIFs, where present
  stage_heldout/<job>/stage_NN/ held-out evaluation after each stage (seq10 and pool10)

<job> is <task>_<mode>, e.g. pen_po_pool10_L3PG.

Reading results in Python

import glob, json
root = "experiment_results/opus55_po_pool10_L3PG_general_pen_r1"
job = "pen_po_pool10_L3PG"
summary = json.load(open(f"{root}/heldout/{job}/evaluation.json"))["summary"]
print(summary["successful_episodes"], "/", summary["planned_episodes"])

state = json.load(open(glob.glob(f"{root}/runs/{job}/runs/repetition_001/case-*/attempt_01/fewshot_state.json")[-1]))
print("requests", state["provider_requests"], "search rollouts", state["search_rollouts"])
print("training positions passed per submission", [len(s.get("passed_case_ids") or []) for s in state["submissions"]])

for path in sorted(glob.glob(f"{root}/heldout/{job}/episodes/*/result.json")):
    r = json.load(open(path))
    print(r["case_id"], r["seed"], r["success"])

# one held-out trajectory: each line has the action and the observation before/after it
steps = [json.loads(l) for l in open(f"{root}/heldout/{job}/episodes/heldout-01/trajectory.jsonl")]
print(len(steps), steps[0]["action"][:4], steps[0]["after"]["state"]["values"][:4])

Notes

• The account e-mail address in the Claude CLI session transcripts is replaced by <redacted-email>; nothing else was changed. Absolute paths of the original machine appear in logs and configs.
• Reference papers under papers/ are text extracts of published papers, included as the agent saw them.
• Results are one run per cell unless the campaign name has _r2/_r3; small differences are within run-to-run variation.