Case study · Research prototype · one supported strategy pack
A solo-play simulator for one competitive Commander strategy, built to state what it cannot see.
Competitive Commander (cEDH) players argue about which cards make a deck faster, but rarely with numbers that have stated limits. A simulator that plays one deck alone, often called goldfishing, can measure how quickly a declared combo state comes together. But a single fast number is easily misread as deck quality. This project measures one thing for one deck, Kinnan, Bonder Prodigy: the turn a declared board state is assembled with nobody interfering. It also reports which cards the model cannot see.
The project work spans an isolated C++ simulation core, authored strategy mechanics, a read-only data exporter, a versioned service boundary and reproducibility checks. Dylan owns the repository. The reviewed records do not individually attribute those implementations or design decisions to him, so manual authorship is not claimed.
The cloud-alignment plan (work items W1–W9, 2026-09-04) is written as instructions for a coding agent, and the HTTP service, Linux image and CI changes followed it. Earlier phases have no file-level attribution, and the tools used are not recorded.
Decision 1
The result field is goldfish_turns_to_assembly. Comparison columns describe assembly differences, and patterns are named for board states rather than wins. The documented Kinnan report groups 31 inert cards by reason and discloses a known misclassification. The useful result is accompanied by the modeling gaps needed to interpret it; an apparently precise number is not presented as a general card ranking.
Decision 2
A request must name an installed strategy pack and a matching commander. The simulator recomputes the deck hash itself and rejects a mismatch, and an unsupported pack is refused with a distinct error code. Nothing falls back to Kinnan behavior. This keeps stored results trustworthy, but the service supports one strategy until someone authors another pack.
1.41% / 20.86% / 76.82%
Historical post-R3 model record in SIM_PLAN §16.11: 60,000 games using list A, a September 1 decklist snapshot. These numbers belong to that model revision, not earlier ablation experiments or current deployed behavior. Assembly while playing alone is not win rate or deck strength.
0.68 s / 297.61 s
Recorded GitHub Actions Ubuntu workload on two threads, run 33906314023: 20,000 games for the interactive run; 98 ablations at 30,000 games each for the sweep. Different workloads, not a speedup comparison. This motivated separate interactive and batch deployments; it is not a concurrent-service latency guarantee.