NEXT context20 saved points · 120 calls · offline replay| Judgment | Off | On |
|---|---|---|
| Next milestone | 35/40 | 40/40 |
| Good tactic | 37/80 | 68/80 |
| Repeated failure | 17/40 | 2/40 |
| Wandering | 7/40 | 1/40 |
NEXT off37/80
NEXT on68/80
Select a moment to open its complete timelapse.
NEXT context20 saved points · 120 calls · offline replay
Decision modelsLive runs| Model | Result | s/step |
|---|---|---|
| Opus 5.5 | All passes + both portal entries | 5–8 |
| GPT-6-Sol | 0 → 8/10 frame blocks; 3 continuations | ~19 |
| Fable 5.1 | Full kit; no portal progress | — |
| Sonnet 5.5 | Early Commander / Direct v1; 0/20 stage passes* | — |
*With Opus escalation. Builds and starting states differ.
Motor policiesOffline only| Controller | Probe | Result |
|---|---|---|
| VPT | Baseline + recurrent replay | Evaluated; no live deployment |
| STEVE-1 + MineCLIP | “Chop a tree” | Aimed at logs 3–4× faster than random; found them no more often |
| ROCKET-3 + SAM 2 | Aiming | 16/39 targets reached; 4 held attacks |
Run cost200–800decisions / attempt
30steps / chapter limit
150ktokens / chapter limit
~1,555image tokens / 4-view scan
Game ticks freeze during inference. Wall time continues.
Practice build · oracle state excluded from agent inputNo full completion recorded