Spatial context and survival policies

Last Hearth asks the model to choose one available role action for one villager. Code supplies observations, selects a legal target for the chosen behavior, navigates, and resolves combat. The model does not output XY coordinates, invent entity IDs, or generate an executable plan.

flowchart LR
  W[Current game state] --> O[Role-specific observations]
  P[Editable role policy] --> Q[Qwen in WebLLM]
  O --> Q
  Q --> L[Scores for available labels]
  L --> A[Fixed role action]
  A --> N[Navigation and combat]
  N --> W

Make geometry explicit

Rounded distances and compass bearings describe spatial relationships. Time to home and time until starvation help the model decide to eat before the return journey becomes fatal. Collectors see route danger and round-trip estimates. Fighters see allies/buildings already under attack. Builders see damage and remaining defense sites and wounded allies. Unrelated details are omitted to reduce prefill.

Click a villager to see its exact current context. The information comes from simulation state, without a vision call or fog-of-war mechanic. The nearest-orc subset limits prompt size; it does not model realistic perception.

reaches_me_seconds estimates a direct approach at the orc's current speed. It does not predict that an orc will change targets or account for every detour. Route risk measures proximity to current orcs; movement after observation can invalidate it. Safer routing is not invulnerability. Taunts work only within the game's specified range.

Put dependent behavior inside one action

Independent attack, target, and taunt fields could form an incoherent combination. Instead, defend selects that behavior as a whole. The game chooses a threatening orc, approaches, taunts, and attacks. Similarly, forage_safe includes harvesting and the return trip. These are defined game mechanics, not an invisible policy replacing a bad model choice.

Basic needs are a visible game rule: at 28 hunger, a villager temporarily stops its job, travels to real food, eats, and resumes the requested work. Shared meals and builder treatments consume supplies. The inspector distinguishes the model's requested job from automatic self-care; meals and single-option rules never count as AI ticks. Zero hunger still kills, and an empty world cannot feed anyone.

The candidate set excludes nonexistent work: no repair without a damaged standing building, no healing without a patient and stored food, no defense without a live orc. Relaxation is available for fatigue, hunger, wounds, an unfinished break, or when no productive job remains. Stamina drains during work and recovers at home; character-specific break thresholds are explicit game rules. A healthy, fed, rested builder returns to useful available work. A stale result is rejected if its job has become unavailable or the actor is now eating. Dead actors receive no commands. These constraints guarantee a legal action, not a wise one.

Tune with evidence

Keep seed and simulation speed fixed, then change one role policy at a time. Maintaining food reserves, interrupting training to protect allies, and balancing repair against new defenses are useful priorities. Independent choices can still conflict: six sensible eating choices may empty the pantry faster than two collectors replenish it.

The policy fixtures cover urgent hunger, peaceful work, threatened allies/buildings, and construction. They are small scenario checks, not a representative intelligence benchmark. The scripted balance runs exercise multiple seeds using explicit rules. They do not measure Qwen or establish globally optimal prompts. Seeds fix random events, while model arrival times and frame cadence can also alter gameplay.

One browser engine processes one actor at a time. The next actor receives fresh observations after the previous call completes. Round-robin scheduling prevents eligible roster members from being skipped. Actors busy with automatic meals are not scored, and single-choice states use an explicit rule without a model call. Pauses, resets, and instruction edits invalidate pending results. A result can still describe a slightly older world by the time inference completes; this delay is part of the challenge and appears in the timing counters.

The formatter spells out the direction of the food meter, empty repair lists, current attack targets, and travel estimates in plain language. The inspector shows those same facts. Role instructions remain editable. Experiments with extra examples, explicit thresholds, semantic labels, and a 2B model did not establish a reliable survival improvement. The earlier 0.8B policies missed important actions; recorded results include failures. Those scene results predate the current needs and healing rules. The new character prompts give Mira a cautious style, Bram a preference for leisure, Aldric a protective role, Sable an appetite for training, Tomas a focus on construction, and Nell a focus on treatment. They are prompt instructions, not guaranteed behavior or an optimal controller.

Input truth and output structure differ

Last Hearth reads numeric state directly. An image-based application can first ask a vision model to describe a scene and then use JEVfire to score finite actions. Perception errors remain possible even when output fields are exact. Image example and current runtime limits.