AI systems
Chronicles
Private repository — the publicly agreed part is shownA text game where code rolls the dice, not the model
- 500
- stubbed runs
- ≈$0.24
- cost per run
- 46
- tests
Problem
Let an LLM decide whether the player's action succeeded and it will bend the outcome toward the prettier scene. The sense of an honest roll — the only thing separating this from AI Dungeon — evaporates in the first twenty minutes.
System
Before the roll the model names the relevant skill and a difficulty; after the roll it narrates an outcome already decided. All the math — the roll, resources, death, card selection — lives in a deterministic engine with zero dependencies and zero API calls. Difficulty is derived in two steps: real-world plausibility on an anchored scale, then a correction for the generated world's explicit laws. Every option must carry a difficulty rationale — the validator rejects options without one.
Role
Solo build: engine, LLM layer, simulation harness.
Outcome
The engine is debugged without a single billed call: 500 stubbed runs caught four balance defects, including one where succeeding at checks made a skilled player die sooner than a random one. That defect only shows up when two strategies are compared — so the comparison is pinned by a test. 46 tests; a run costs ≈$0.24 against a $0.28 budget.