AceWAM Press Button GDN-surprise router — step 6000
Public evaluation bundle for the adaptive Full/Gist memory model trained on
RMBench press_button.
The router has no fixed Full-page budget. The most recent four memory pages are
kept Full; when an older page expires from that window, its learned causal Gated
Delta surprise score is compared with the running causal median plus one MAD.
The page is then stored exclusively as either 120 Full tokens or 8 Gist tokens.
Files
step_006000.pt: inference checkpoint.config.yaml: fully resolved training configuration.dataset_stats.json: action and proprio normalization statistics.manifest.json: immutable job, Git, size, and SHA-256 provenance.
Provenance
- Training job:
pt-exc4qkxv
- Training run:
press_button_gdn_surprise_a100_1x8_152e8fc_r4
- Training code:
sgres-paw/AceWAM@152e8fc204279d87e3c1b798852a46b9b3002fb8
- Evaluation instrumentation:
sgres-paw/AceWAM@d4d6fca6bd7ef793fecf4daf1eb18d76e4259ded
- Checkpoint step:
6000
This bundle contains model weights and evaluation metadata only. RMBench assets
and the Wan base-model dependencies must be supplied separately.
