TerranSoul / Zork bench / Claude Opus 4.8 ← TaughtLocalLLM ← Back to main

Claude Opus 4.8 — real Zork I run

Opus 4.8 playing Jericho zork1.z5 move-by-move (real game responses), using the ZorkGPT agent logic + the TerranSoul brain as memory. Real runs only — no engine auto-solver.

350

Recall-based real run: Opus reproduced the Zork solution it knows from training, played move-by-move on Jericho (real game responses, the winning "barrow" ending, 395 moves). This is recall, not from-scratch reasoning — Opus's cold-reactive (no-recall) score is 50, which is below deepseek's scaffolded 94. So Opus leads deepseek via recall, not raw in-context reasoning; a like-for-like Opus-in-the-ZorkGPT-scaffold run isn't possible here (no Opus API). Genuine agency verified on the opening (diagnosed the chimney 3-item limit via inventory, score 45).

Honest baseline (real runs only)

RunScoreWhat it is
Opus 4.8 — cold reactive (no recall)50genuine in-context play with zero Zork recall
Opus 4.8 — recall-based, move-by-move350reproduces the solution it knows; real Jericho run, the winning game
ZorkGPT (deepseek-v3.2 + 27B)94–115frontier baseline to surpass
TaughtLocalLLM (gemma4:e4b)177weak 4B + taught solution context

"Opus leads deepseek" holds when Opus leverages its Zork training (recall) move-by-move — the same basis deepseek uses. The cold-reactive floor (no recall) is 50.

Replay (genuine opening)

> north / east / open window / enter window     house            score 10
> west / take lantern / take sword
> move rug / open trap door / turn on lantern
> down                                           Cellar           score 35  (+25)
> north / kill troll with sword                  troll dies
> south / south / east                           Gallery
> take painting                                  painting         score 39  (+4)
...continuing

Full per-turn transcript + score curve recorded with the completed run.

How the brain works together

One coherent runtime loop

Client request a question arrives Connecting nerve shared access channel Focus & triage what matters now Hybrid recall fuse signals, rerank to a short grounded context Language model core answers from context (local or cloud) Three-tier memory recent, durable, archived + knowledge graph Outcome observed what actually happened Self-improvement loop classify outcome, reinforce or quarantine, update memory Brain’s sources connectors repo brains peers over the nerve need context pull grounded context answer observe write back feed

Many specialised parts, one coherent loop.