docs/reproduction.md

# Reproduction recipe

How the datasets in [this repo](../README.md) were recorded from public
materials. Scripts live in the GitHub repo — record with
[`examples/unity_collection`](https://github.com/ccnets-team/causal-gpt-rl/tree/main/examples/unity_collection)
and package with
[`collection/build_minari.py`](https://github.com/ccnets-team/causal-gpt-rl/blob/main/collection/build_minari.py).
For each ONNX-driven environment, get the matching model-removed build and stock
policy from the
[envs repo](https://huggingface.co/datasets/ccnets/causal-gpt-rl-unity-envs),
record each agent's per-step `(observation, action, reward, termination)`, and
preserve the action/observation structure declared by that build.

## Format & provenance

All environments, datasets, and stock behavior policies use **ML-Agents
`release_23`**; machine-readable provenance is in `release_manifest.json`. Every
dataset is stored at `<name>/<tier>/data/main_data.hdf5` with a sibling
`metadata.json` (`minari_version` 0.5.3).

**Action labels and decision periods.** An ML-Agents agent requests a decision
only every *k*-th step; Unity repeats that decision through the gap, and the agent
does not appear in `DecisionSteps` again until the next one. For those in-between
steps, record the action Unity is still running — carry the agent's last decision
forward. A recorder that instead stores whatever its policy wrapper emits for an
agent with no observation that step will label most transitions with a
placeholder; see the dataset card's data-quality notice.

**Ego-agent schema (SoccerTwos, DungeonEscape).** These use
`observations["agents"]["agent_0"]` / `actions["agents"]["agent_0"]`. Each
physical agent is stored as one ego-centric episode because the Unity examples run
one shared decentralized policy independently per agent; the wrapper exposes the
multi-agent structure without granting the policy privileged access to teammates'
private observations. Reconstruct a match/group by grouping the linked episodes on
`match_id`; **split by `match_id`, never by individual ego episode.**

**Quality tiers.** `medium`/`simple` degrade the stock policy (seed 2310000):
continuous scenes add per-episode Gaussian action noise from a calibrated range;
the discrete goal games sample the policy's own logits via `softmax(logits/T)` and
are scored by step-reward (return/length, since their episode return is bimodal) —
PushBlock and Pyramids each sampling `T` **per episode** from a calibrated range
(each episode its own skill level); DungeonEscape (cooperative) samples a
softmax temperature `T` **per group per match** from a calibrated range — all three
group agents share the match's `T`; competitive self-play SoccerTwos draws a per-team
softmax temperature `T` independently per match from a (wider) range — the two teams
draw independently, so each match pairs two skill levels (a random gap = league play).
Exact values, anchors, and calibration notes are in the
[dataset card](../README.md)'s "Quality ladders" section.

## SoccerTwos

1. Build the release_23 SoccerTwos scene with both teams' model assets removed so
   both behavior groups are controlled externally.
2. Run the stock `SoccerTwos.onnx` for both teams and record both sides. The
   original ONNX declares two nonexistent Sentis metadata graph outputs; for ONNX
   Runtime, remove only those dangling output declarations (the action graph is
   unchanged).
3. Store the sum of per-agent `reward` and `group_reward` (goals arrive via
   `group_reward`).
4. Preserve `Tuple(Box(264), Box(72))` observations and `MultiDiscrete([3, 3, 3])`
   actions.
5. Link the four decentralized episodes per field with `match_id`; finish all
   in-flight matches after crossing one million transitions.

Expert validation: 1,001,676 transitions, 14,692 terminated episodes, 3,673
matches, 0 truncations, every match has equal linked lengths with two wins / two
losses. Mean agent ep length 68.18, mean return −0.034089.

Tiers: `medium`/`simple` draw a **per-team** softmax temperature `T ~ U(1.3, 2.6)` /
`U(2.1, 3.4)` (each team draws once per match, independently of its opponent, shared by
its two agents — a per-match skill gap = cross-checkpoint league play; `collect.py
--team-temperature-range <lo,hi>`). Self-play keeps returns balanced (both tiers ≈ 50/50
win-loss), so tier skill is set by the ego team's own drawn `T` and calibrated to
side-swapped match-score norm ≈0.80 / ≈0.60 (expert T=1 → 1.0, random T=25 → 0). medium:
12,940 ego ep / 1,002,368 transitions / 3,235 matches; simple: 11,820 / 1,001,564 / 2,955.

## DungeonEscape

1. Use the stock `DungeonEscape.onnx` shared POCA policy with the model removed so
   each agent is driven externally.
2. Record each physical agent independently; preserve its three local sensor leaves
   (rigid-body 10, stacked ray 360, has-key flag 1) and single `Discrete(7)` action.
3. Relabel each physical agent as ego path `agents.agent_0`. Three episodes sharing
   one `match_id` reconstruct the cooperative group (distinct local trajectories,
   not copies).
4. Sum `reward` and `group_reward`. Determine success only at group level — an
   agent that dies early may not receive a later positive group reward.
5. Split by `match_id`.

Expert validation: 1,007,206 transitions, 35,505 ego episodes, 11,835 groups
(11,366 successes / 469 failures), 0 shape/range errors. Every match links exactly
three episodes.

Tiers: `medium`/`simple` apply a **per-group** softmax temperature `T ~ U(1.5, 3.0)` /
`U(3.0, 4.0)` (drawn once per cooperative group per match and shared by its three
agents), calibrated to step-reward norm 0.80 / 0.61 against the expert step-reward
anchor 0.019355 (random ≈0.00015). medium: 28,314 ego ep / 1,012,457 transitions;
simple: 22,542 / 1,001,594.

## PushBlock

1. Use the model-removed PushBlock build and drive all 32 agents with the stock
   `PushBlock.onnx`. This is the original discrete PushBlock, not the hybrid
   PushBlockWithInput.
2. Preserve the two `105`-d sensors and the single `Discrete(7)` action branch.
3. Stop at one million transitions and package each physical-agent episode
   independently.

Expert validation: 53,039 episodes (52,946 terminated / 93 truncated at the
recording boundary), mean return 4.972780, mean ep length 18.85.

## Crawler, Pyramids, Worm, Walker, 3DBallHard

Use each model-removed build with its matching stock ONNX; record each physical
scene agent independently and stop at one million transitions. Schemas come
directly from the live ML-Agents behavior specs.

- **Crawler** — 1,043 episodes, 1,000,000 transitions, mean return 2575.8405, mean
  length 958.77 (76 terminated / 967 truncated). `Tuple(Box(126), Box(32))`,
  `Box(20)`.
- **Pyramids** — 5,392 episodes, 1,000,000 transitions, mean return 1.797730, mean
  length 185.46 (5,344 terminated / 48 truncated). `Tuple(Box(56)×3, Box(4))`,
  `Discrete(5)`.
- **Worm** — 1,000 episodes, 1,000,000 transitions, mean return 1044.106317, mean
  length 1000.00 (0 terminated / 1,000 truncated). `Box(64)`, `Box(9)`.
- **Walker** — 1,464 episodes, 1,000,000 transitions, mean return 1354.565443, mean
  length 683.06 (768 terminated / 696 truncated). `Box(243)`, `Box(39)`.
- **3DBallHard** — 1,011 episodes, 1,000,008 transitions, mean return 98.909511,
  mean length 989.13 (3 terminated / 1,008 truncated). `Tuple(Box(27), Box(18))`,
  `Box(2)`.

The recipe ends at the Minari dataset — a portable, env-less trajectory set.