# Episode analysis and resource evolution

Model: GPT-6 Astra / low, standard tier, fast disabled. Native successes: Kinex 1/5; Codex 3/5. The four Kinex failures reached an additional 100-model-turn cap. Native evidence is valid for all ten episodes, but the campaign is not execution-valid.

## Per-episode evidence

### Codex E1: Stopped just below the threshold

The maximum target-door angle was about 80.27 degrees, below the native 81-degree threshold; it relaxed to 79.67 degrees at termination. The agent voluntarily finalized with 2614 control steps remaining.

Termination: **stopped**. Control steps: **2386/5000**.

Created tools/panda.py and memos/robocasa.md. The visual-angle explanation in the memo disagrees with the native-state audit.

[Detailed evidence](episodes/codex/episode_01/analysis.md) · [Session](episodes/codex/episode_01/session.txt) · [Tool activity](episodes/codex/episode_01/tool-activity.json)

### Codex E2: Native success

The target fridge door passed the native threshold. The episode reused the prior panda.py tool and memo.

Termination: **success**. Control steps: **794/5000**.

Reused panda.py and updated the memo. The claim that one leaf can suffice for a double-door fridge was an unsafe generalization across fixture types.

[Detailed evidence](episodes/codex/episode_02/analysis.md) · [Session](episodes/codex/episode_02/session.txt) · [Tool activity](episodes/codex/episode_02/tool-activity.json)

### Codex E3: Second fridge door omitted; voluntary stop

The left French-door leaf opened to about 89.96 degrees, but the right target leaf stayed closed throughout. The agent voluntarily finalized with 379 control steps remaining.

Termination: **stopped**. Control steps: **4621/5000**.

Added an unreachable-pose guard to panda.py and created pivot.py with lost-grip/tracking-error stops. Updated the memo, but retained a qualified version of the earlier one-leaf claim.

[Detailed evidence](episodes/codex/episode_03/analysis.md) · [Session](episodes/codex/episode_03/session.txt) · [Tool activity](episodes/codex/episode_03/tool-activity.json)

### Codex E4: Native success with reused tools

The agent reused panda.py and the previous episode's pivot.py, adjusted the grasp and opening motion, and reached native success. Tool reuse is observed; its causal benefit was not isolated.

Termination: **success**. Control steps: **989/5000**.

Read prior tools and memo; executed pivot.py twice. Updated the memo with the successful procedure; tool files were unchanged.

[Detailed evidence](episodes/codex/episode_04/analysis.md) · [Session](episodes/codex/episode_04/session.txt) · [Tool activity](episodes/codex/episode_04/tool-activity.json)

### Codex E5: Native success after compartment recovery

The run opened the freezer and then the target fridge door, finally reaching native success at 4888 control steps. The agent updated its notes, but no sixth episode tested those final updates.

Termination: **success**. Control steps: **4888/5000**.

Reused prior tools and updated the memo with right-side opening, rotation-composition and base-motion lessons. No subsequent episode validates these final notes.

[Detailed evidence](episodes/codex/episode_05/analysis.md) · [Session](episodes/codex/episode_05/session.txt) · [Tool activity](episodes/codex/episode_05/tool-activity.json)

### Kinex E1: Partial opening; model-loop cap

The target fridge door reached only 71.95 degrees. Repeated approach, IK and regrasp attempts did not complete the opening before the 100-turn harness cap.

Termination: **100 model-turn limit**. Control steps: **1336/5000**.

Created base_step and hold_steps. No memo was saved before the model-turn cap.

[Detailed evidence](episodes/kinex/episode_01/analysis.md) · [Session](episodes/kinex/episode_01/session.txt) · [Tool activity](episodes/kinex/episode_01/tool-activity.json)

### Kinex E2: Native success

The target fridge door passed the native opening threshold. A subsequent device_stop was rejected because the episode had already finished; this cleanup warning does not invalidate native success.

Termination: **success**. Control steps: **1949/5000**.

Created tcp_move and open-fridge.md. These resources were inherited and used in episodes 3–5.

[Detailed evidence](episodes/kinex/episode_02/analysis.md) · [Session](episodes/kinex/episode_02/session.txt) · [Tool activity](episodes/kinex/episode_02/tool-activity.json)

### Kinex E3: Second fridge door omitted

The left French-door leaf reached the threshold, but the right target leaf stayed closed throughout. The public goal used the plural "doors". The 100-turn cap then terminated the agent.

Termination: **100 model-turn limit**. Control steps: **3763/5000**.

Read the saved memo and tool; called tcp_move 33 times. No learned file changed relative to episode 2.

[Detailed evidence](episodes/kinex/episode_03/analysis.md) · [Session](episodes/kinex/episode_03/session.txt) · [Tool activity](episodes/kinex/episode_03/tool-activity.json)

### Kinex E4: Insufficient opening; control difficulties

The target fridge door finished near 64.99 degrees. Repeated IK failures, joint-limit difficulties and grasp recovery consumed model turns before the 100-turn cap.

Termination: **100 model-turn limit**. Control steps: **2226/5000**.

Read the saved memo; called tcp_move 32 times and modified its implementation. The memo remained unchanged.

[Detailed evidence](episodes/kinex/episode_04/analysis.md) · [Session](episodes/kinex/episode_04/session.txt) · [Tool activity](episodes/kinex/episode_04/tool-activity.json)

### Kinex E5: Wrong compartment opened

The freezer door opened to about 81.32 degrees, while the target fridge door remained essentially closed. The agent then reached the 100-turn cap.

Termination: **100 model-turn limit**. Control steps: **3023/5000**.

Read the saved memo; called tcp_move 32 times. No learned file changed relative to episode 4.

[Detailed evidence](episodes/kinex/episode_05/analysis.md) · [Session](episodes/kinex/episode_05/session.txt) · [Tool activity](episodes/kinex/episode_05/tool-activity.json)

## Limits of causal attribution

The model-turn cap is the stopping mechanism, not a complete physical failure explanation. Kinex E3 omitted the second door and E5 opened the wrong compartment; removing the cap alone is not demonstrated to solve either. Codex E1 and E3 voluntarily stopped with remaining control budget. Tool errors also occur in successful episodes and are not automatically causes of failure.

No separate Kinex reflect skill execution occurred. The episode-2 Kinex memo was unchanged in episodes 3–5; episode 4 modified tcp_move during execution. Codex updated its memo every episode, and the episode-3 pivot tool was used in later episodes. Some notes incorrectly generalized one fixture's success to a different fridge type. Artifact creation, inheritance, actual use and causal benefit are distinct; only the first three are evidenced here. Paired held-out-state ablations would be needed to measure learning benefit.