Re-encode the replay for the page widget

Replace the replay with the chimera preview

Add a preview video

Publish simple-only Walker2d step 28,800

fix: re-measure Hopper on the protocol runtime and correct the KV sweep

Publish simple-only Hopper step 94,400

Publish simple-only Hopper step 40,400

Update Walker2d simple-only policy

docs: remeasure the kv=1000 column on the protocol runtime

docs: add kv=1000 column to the KV retention sweep

docs: report 100-episode KV8/32/128 sweep

fix: correct Humanoid KV64 score after shared-KV correction

fix: refresh MuJoCo scores after shared-KV correction

docs: say what a token is before drawing one

docs: drop what the Files tab and the code repo already say

docs: put installing next to the versions that reproduce the table

docs: stop steering readers back to 1x

docs: put the table's own legend next to the table

docs: stop calling the trained window a limit

Refresh Ant-v5 with step 99,600 bundle

Refresh Hopper-v5 with step 98,000 bundle

Lead the KV retention sweep with the long-context figure

Add the long-context extrapolation figure

Refresh HalfCheetah-v5 with 0.16.0 step 100,000 bundle

Refresh Hopper-v5 with step 78,400 bundle

Refresh Walker2d-v5 with step 44,800 bundle

docs: size the rollout picture to the model-card column

docs: show the environment transition in the rollout picture

card: fix the KV retention boundary (trained window, not position window)

assets: give the rollout diagram an intrinsic size so <img> embeds render

docs: lead with the rollout action-generation diagram

assets: add the rollout action-generation diagram

card: point the evaluation protocol at the reproduction script

Drop v1 bundle references

Restore Ant/Walker2d rows, add the fold figure, narrow 'no critic'

Shorten KV1000 table header

Remove KV sweep trend column

Fix KV retention table rendering

Correct Ant and Walker2d KV1000 normalization

Document KV1000 retention evaluation

Update Walker2d and Ant policies for v0.16

Update Walker2d policy to stable step 34400

docs: clarify context_length vs kv_cache_max_len in the retention sweep

Link the public W&B project for the MuJoCo runs

Update Walker2d and Ant policies

Publish Pusher-v5 medium-only policy

Publish Swimmer-v5 medium-only policy

Remove premature W&B link from model card

docs: ship the PolyForm Noncommercial 1.0.0 text in the repo

Update HumanoidStandup-v5 to pc4 320k