SemIf HTTP API and rocm flake a3fc1b5fcb0bd481dcc8579e5a2e4182ac68e75e 2026-09-24T14:49:03Z 2026-09-24T20:53:09Z James Eversole james@eversole.co commit a3fc1b5fcb0bd481dcc8579e5a2e4182ac68e75e Author: James Eversole <james@eversole.co> Date: Thu, 24 Sep 2026 09:49:03 -0500 semif-api: HTTP API around SemIf + llama backend FastAPI wrapper (SemIf consumed as a library, no fork) exposing its semantic decision readout over HTTP: /decide and /decide-batch read the softmax over declared options from logits (torch ROCm backend) or from an OpenAI-compatible llama-server hosting a GGUF. The llama backend adds /plan, a reasoning chat completion that turns an action transcript into rules, plus runtime model selection. The browser UI (served same-origin) includes four demos: platformer and puzzle room, each in ruled and self-learning variants. The self-learning loops are the point of the repo: the model plays from neutral observations alone, and on failure a universal planner prompt rewrites its rules from the transcript of what happened. Packaging: nix flake (upstream SemIf as an input, ROCm dev shell for gfx1151, NixOS module with systemd service), offline test suites for the game prompts and the backend contract, validation script, docs.