AI-ported games: does it build?

Results by reported outcome

Each table is one group of the source list, by the outcome the project reported. Reproducible means built or static_web. Of the 267 projects listed, 13 have no repository and were not checked.

Build files found below the root: 66. Submodule clones that fell back to a plain clone: 2. Retries: sdl3 6, clang 3.

Why a build did not finish

Every result other than built or static_web, broken down by the reason the harness recorded. The reason is mined from the build log with word lists, so read it as a signal.

What was missing

The names the classifier pulled out of failed logs: packages, headers, tools and files the build asked for and did not find. At most six per repository, top 15 across the set.

Model family, Works group

The model each project named on the source list, grouped into families, crossed with the build result. Claude covers any mention of Claude or Anthropic; OpenAI covers GPT, Codex or OpenAI; a project that names both counts under Claude. This is the project's own description of its tooling.

Last commit

Age of the most recent commit on each repository's default branch at the time of the check, counted from October 29, 2026.

Every repository

One row per repository, sorted by reported outcome then id. The title links to the repository. Detail is the matched log phrase, cut to 80 characters.

Method

Result values

Reason values

Limitations

• A build file three levels down is found; one deeper is not. Build files under vendor, third_party, tests and similar directories are ignored on purpose.
• The shallowest directory wins. A root build.sh is chosen over a src/CMakeLists.txt, on the theory that the repository's own script knows how to drive its own tree.
• Only Ubuntu packages and SDL3 from source are supplied. Anything else a project needs (a console SDK, a Windows toolchain, Metal, the original game) shows up as needs_toolchain, missing_toolchain or missing_assets, which is the expected result for most decompilations.
• Retries are limited to two cases, SDL3 and Clang. A project that needs a different compiler version, a nightly Rust or a specific Python is reported as it failed, not fixed.
• The reason patterns are word lists. They over-match now and then (a stray "retail" or "wasm" in a log) and under-match when the real error scrolled past the 600-line log tail, which lands in unclassified. A second pass over the kept tails can only move a row from unclassified or compile_error to a more specific reason, never the other way, and such rows carry reason_source and reason_runner.
• The asset check is a phrase scan of the README and build docs and over-matches short words like "rom" and "iso"; only the strict phrases are counted.
• Nothing here says whether the built program runs, renders or matches the original.
• One repository on the partial list did not complete because its runner was shut down twice; it is absent rather than guessed.