Skip to content

joint_s4 — cocarry_partners@v2

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