Public showcase · Simulation & Evaluation
Cuttle LLM Tournament.
LLM gameplay experiments comparing Claude, GPT/OpenRouter, local models, heuristics, and search strategies inside the same Cuttle engine.
Metric · 01
Opus result and model comparison notes referenced in Git history
01 / Showcase Signal
What this showcase helps explain.
The tournament asks a practical question: when does language-model reasoning beat rules, heuristics, or search in a hidden-information card game?
- Opus result and model comparison notes referenced in Git history
02 / Architecture
How the system is put together.
Cuttle LLM Tournament starts with a concrete game ai strategy problem, routes it through LLMs, Simulation, Evals capabilities, and produces an artifact a builder can inspect.
The useful pattern is the boundary between automation and proof: inputs are captured, the AI-assisted step is constrained, and the result is checked through screenshots, source links, tests, reports, or public product surfaces.
03 / Walkthrough
The reproducible pattern.
01
Start from the source artifact
Collect the public URL, local output, report, or dataset that proves the game ai strategy workflow exists and can be inspected.
Input artifact: https://github.com/elawless/cuttle-simulation
Privacy check: Redact API keys and any raw provider responses if included in logs.02
Apply the LLMs layer
Use llms as the main transformation layer, then keep the intermediate result visible enough for review.
03
Publish the proof surface
Ship a screenshot, diagram, report, dashboard, or link that makes the outcome understandable without asking the visitor to trust the description.
04 / Stack
What it leans on technically.
| Tool | Version | Role | Why this tool |
|---|---|---|---|
| LLMs | Current | LLMs | Adds language understanding, classification, extraction, or reasoning where deterministic code is too brittle. |
| Simulation | Current | Simulation | Creates repeatable trials and counterfactuals for comparing decisions safely. |
| Evals | Current | Evals | Keeps generated or automated work accountable through tests, comparisons, and review artifacts. |
| MCTS | Current | MCTS | Searches decision trees through simulation so strategy can be measured rather than guessed. |
05 / Prompts
Copyable prompt surfaces.
Inspection prompt
SystemYou explain AI workflow evidence clearly without inventing private implementation details.
User templateProject: Cuttle LLM Tournament Evidence: <artifact> Explain the input, AI step, validation, and public output in plain English.
06 / Evidence
Artifacts you can inspect.
github pr LLM tournament infrastructure commithttps://github.com/elawless/cuttle-simulation/commit/42f7d858aa129cc561073a91efb92a06877e81ec Available
07 / Privacy & Evidence Boundary
What the public material does and does not expose.
Privacy. Redact API keys and any raw provider responses if included in logs.
08 / Next Step