Cookbook: Run Qwen in the Minecraft LLM-Agent Harness

This cookbook records the practical setup used for the Qwen lanes in the natural-village model comparison. It is intentionally small: the goal is to make the Qwen part reproducible enough for another developer to inspect or adapt.

What This Runs

The target lane is a 30-cycle Minecraft run where an LLM-backed actor starts near a natural village and tries to make grounded material progress:

• collect logs;
• craft planks and sticks;
• create or place a crafting table;
• recover from blockers;
• leave runtime artifacts and screenshots for review.

The Qwen comparison used:

• provider: modelscope-api
• models: Qwen-Ambassador/Qwen3.7-Plus, Qwen-Ambassador/Qwen3.7-Max
• scenario: natural-village-spawn-v1
• seed: 4167799982467607063
• visual profile: report
• server version for visual evidence: 1.21.4

Preflight

Run provider preflight before spending quota. The exact preflight artifact used for the comparison is stored at:

project-docs/experiments/curated/2026-06-29/goal-oriented-natural-village-30cycle-qwen/preflight/qwen-30cycle-preflight.json

The run should be treated as provider-backed and budget-sensitive. Do not start with a long cycle count until a short provider smoke has passed.

Example Run Shape

The Qwen lanes were produced through the social-cycle CLI with the natural village scenario and report visual profile.

The command shape is:

bun run probe:social-cycle -- \
  --provider modelscope-api \
  --model Qwen-Ambassador/Qwen3.7-Max \
  --actor npc_b \
  --cycles 30 \
  --max-actions-per-cycle 1 \
  --world-scenario natural-village-spawn-v1 \
  --visual-profile report \
  --report ../project-docs/experiments/curated/2026-06-29/goal-oriented-natural-village-30cycle-qwen/reports/qwen-3.7-max.json

Use Qwen-Ambassador/Qwen3.7-Plus for the Plus lane.

What To Inspect

Do not judge the run from model text alone. Inspect these artifacts:

• run report JSON;
• review summary JSON;
• visual evidence audit;
• provider usage ledger;
• inventory deltas;
• verifier status;
• selected first-person and third-person screenshots.

For the final comparison, start with:

project-docs/exports/static/final-natural-village-model-comparison-report-2026-06-30.html
project-docs/experiments/curated/2026-06-30/goal-oriented-natural-village-gemini-30cycle-extension/combined-model-comparison-30cycle-analysis.json

Observed Qwen Result

In this run:

Lane	Verified progress	Blocked	No progress	Provider records	Total tokens
Qwen Plus	12	8	10	38	1.13M
Qwen Max	15	8	7	38	1.09M

Qwen Max produced the strongest material progression among the completed lanes: wooden pickaxe evidence, six cobblestone, and later material continuation.

Qwen Plus produced a stable early chain around logs, planks, sticks, and crafting-table evidence.

Known Caveats

• This is not a general model ranking.
• The harness includes model behavior, action selection, Mineflayer execution, verifier contracts, generated action-skill handling, and provider quota effects.
• Screenshots are review-only evidence. Runtime artifacts own block identity and progress claims.
• Some camera modes can produce misleading cross-section-looking images. Treat those as visual caveats, not world truth.

Next Better Experiment

The next run should narrow the objective:

After placing or locating a crafting table, maintain a safe village-adjacent
work point for continued material work.

That target is better than asking for workbench, shelter, shared storage, and long-run social continuity in a single 30-cycle lane.