Executive summary
int-llm-brainfuck implements a small transformer using only Brainfuck's eight commands. Its published generators produce inference and training programs for a pinned machine with wrapping eight-bit cells. Fixed-point arithmetic, model loading, attention, sampling, gradient computation, and optimizer updates happen inside the generated program. A generic interpreter supplies execution and byte I/O. The completed experiment has recorded exact comparisons with the C oracle, including the full 5,000-step training run.
The generated inference file is 16.40 GB and the trainer is 210.30 GB, using decimal GB (1 GB = 1,000,000,000 bytes). The repository contains the generators and verification records; the generated .bf files are not committed.
What was built
- A 16,395,194,474-byte inference program and a 210,300,828,662-byte trainer,
- A pinned 4,194,304-cell Brainfuck machine, portable interpreter, and generic
- Integer primitives, model loaders, forward/backward computation, Adam,
- Oracle fixtures, staged checks, negative controls, and sealed verification logs.
regenerated from the published source rather than stored in Git.
Linux x86-64 ahead-of-time runner.
gradient clipping, and checkpoint serialization.
How it works
Q16.48 values occupy eight little-endian byte cells. Inference keeps stored weights as signed-int16 F12 codes and widens them for computation. Training retains wide parameters and optimizer state throughout; F12 rounding happens only when exporting the completed checkpoint. The runner may optimize generic Brainfuck operations but may not perform model-aware work or supply answers.
Evidence and validation
The recorded inference gate checks 20 generated names, 122 forward steps, and both output and raw-logit hashes. A deliberately corrupted checkpoint preserves the names but changes the logit hash, testing that the gate detects hidden drift.
The September 18, 2026 training record reports exact agreement on 5,000 loss lines, 20 samples, the 115,576-byte wide checkpoint, and the F12 checkpoint. An independent conversion also matches the F12 export. The full run used the generic native runner and took 409:29:53 including compilation. These are committed project records; this portfolio refresh did not repeat the long run.
Boundaries
The result is limited to the pinned small model and machine semantics. It does not establish useful throughput or general language-model quality. The full portable-interpreter training gate was not rerun. Generated programs require substantial disk space, and the training dataset is fetched separately.
Repository map and first run
Start with the small artifact and layout checks:
make bootstrapRead README.md, SPEC.md, and the two reference/phase*-verification-results.log files before generating or running the large programs. The logs retain earlier failed attempts alongside the completed gates.