Quest-Breaker: Narrative Deadlock & Inconsistency Engine for Branching Games
Users: narrative designers and QA teams on branching RPGs and story games. Problem: a writer lets the player kill a character in Act 2, and a quest in Act 5 assumes that character is alive. The player hits an impossible scene or gets soft-locked. There are millions of playthroughs, so manual testing cannot cover them, and existing tools only check explicit flags, not the story text where many contradictions live. Outcome: upload the quest graph as JSON and get, in seconds, every deadlock with the exact failing playthrough and an AI-written in-world fix. Three simulated player types (Murder Hobo, Contrarian, Speedrunner) show who gets soft-locked.
Preliminary score
- Value for the user21.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI25.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing18.0 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility12.0 / 15
Data requirements, running costs and a clear next step.
- Originality8.0 / 10
A meaningful difference from common solutions.
A finalist: the live jury scored it 68.0 / 100 on stage.
What the judges wrote
- Claude Opus 5.5Preliminary84/100
- Strengths
- Your problem is concrete and well chosen: a character killed in Act 2 who is needed in Act 5 is a real, costly bug class, and you correctly note that flag-based checkers cannot see contradictions that live only in story text. Your hybrid design gives each part a clear job. BFS over (node, state) pairs handles exact flag logic, while the LLM reads only edge-covering playthroughs, so the AI is essential for the text-only bugs. Your evaluation is unusually strong for a hackathon: a labeled benchmark with 32 look-alike traps, held-out seeds separate from the prompt-tuning seeds, and a three-way comparison against how flag checkers work today. You also honestly report false alarms, a labeling error, and a case the model still misses.
- To improve
- Your 100% precision and recall come from 16 graphs that your own generator produced, so the next convincing result is a run on real Ink or Yarn scripts, or a public branching game script. Please report any misses you find there. Turn the token counts into a cost per audit for a realistic studio-sized graph of hundreds of quests, and explain how the 500-path cap and edge-cover selection behave at that size. Your materials list both Azure gpt-5.6-sol and Claude as the LLM, so state clearly which model produced the benchmark numbers.
- GPT-6 AstraPreliminary85/100
- Strengths
- Your hybrid architecture gives deterministic state checking and AI narrative analysis distinct, complementary jobs, supported by a concrete API, JSON schema and runnable setup. Your failing-playthrough visualization and proposed fixes make findings actionable for narrative designers. Your held-out synthetic benchmark, clean controls, component comparisons and candid failure reporting provide unusually strong testing evidence for a hackathon.
- To improve
- Your next benchmark should use real Ink/Yarn scripts and independently reviewed labels, measuring designer time saved alongside precision and recall. Your semantic edge-cover sampling and 500-path cap need explicit coverage reporting: visiting every edge does not test every narrative history, so exhaustive detection is not yet supported. Your feasibility case would be stronger with measured latency and monetary cost across graph sizes, plus a consistent description of the model/provider used for the reported results.
The team
Zaka Shamilli
data
Ilqar Farhadov
developer
Mirabbas NabiyevLead
product