# `joint_s4` — cocarry_partners@v2

<!-- Generated by scripts/render_partner_cards.py from chamber.partners.sets + the fingerprint archive — do not edit by hand. -->

**Set.** `cocarry_partners@v2` (task `cocarry@v1`) · **Stratum.** learned · **Split.** public · **Role.** partner_arm

**Identity.** `partner_id` `5118f7828d8ac7b8` · parameter digest `44136fa355b3678a1146ad16f7e8649e94fb4fc21fe77e8310c060f61caaff8a` · `weights_uri` `local://artifacts/e9040cbc3c3b2456_step150000.pt`

## What this partner does

A frozen jointly-trained MAPPO co-carry seat (ADR-011 §Decision as amended): the partner-side actor of a pair checkpoint, loaded with SHA-256 sidecar verification and run deterministic under no-grad (ADR-018/I3 black-box contract). It reads the symmetric full state (own + other proprioception, bar pose, goal) assembled from raw obs leaves — the same interface it trained on. Its cooperative style is whatever joint training produced with ITS ego; cross-play with any other ego is exactly the partner-familiarity question the zoo measures, and the fingerprint below characterises it behaviourally.

## Construction provenance

Built from registry class `frozen_cocarry_joint` (`PartnerSpec(seed=4, checkpoint_step=150000)`). Learned member: behaviour lives in the frozen checkpoint `local://artifacts/e9040cbc3c3b2456_step150000.pt` (committed payload SHA-256 `3107b1bde5201d0350221f527839cda7ee6ac5698c20df6f80eb80453fb5ae18`; the loading wrapper re-verifies the ADR-002 sidecar). **Provenance.** trained jointly with the ego actor of pair checkpoint sha256 3107b1bde5201d0350221f527839cda7ee6ac5698c20df6f80eb80453fb5ae18 (training seed 4, selected step 150000 under the ADR-027 checkpoint-selection rule)

*(no parameter box — a learned member's behaviour is its frozen checkpoint; custody rides on the committed payload SHA-256 above)*

## Behavioural fingerprint

Committed probe suite: seeds [7000, 7001, 7002, 7003] x 5 episodes against the task's reference ego (archive: `spikes/results/partner-fingerprints/cocarry_partners-v2/joint_s4/`, an ADR-028 v3 bundle). Summary statistics of the member's action distribution and the dyad's stress channel — enough to tell members apart without policy access (ADR-018/I3).

| statistic | value |
| --- | --- |
| `action_abs_mean` | 0.1364 |
| `action_abs_p90` | 0.2652 |
| `action_std` | 0.1621 |
| `n_episodes` | 20 |
| `stress_max` | 44.23 |
| `stress_mean` | 33.19 |
| `stress_p50` | 32.57 |
| `stress_p90` | 36.63 |
| `success_rate` | 0.8 |

**Competence floor.** floor probe `fingerprint` — success 0.8 vs committed floor 0.75 → **PASS**