bit-exact

Deterministic, integer-only SLM inference runtime — no floating point on the inference path, so the same model and prompt produce bit-identical output across CPUs, GPUs, compilers, and vendors.

Open, bit-reproducible low-precision number format (b-posit / AI-Posit) for AI inference — 30–50% less memory, with results identical to the bit on any GPU, CPU, or RISC-V. Reference + conformance suite + CORE-ET RTL

Integer-only LLM stack in pure C — run TinyLlama and train a tiny GPT in Q16.48 fixed-point, zero floating point. A reproducible fun experiment.

Pure-Go Opus codec (RFC 6716), zero cgo. A bit-exact port of libopus, differential-tested against the C reference: the decoder passes all 12 RFC conformance vectors, and the CELT encoder produces byte-identical packets.

Visual anatomy of a Q16.48 integer GPT — four self-contained offline HTML pages: weight atlas, bit-exact inference trace, 32-D embedding space, 3-D weight explorer. Every displayed number is exact (BigInt in-browser, verified against the C oracle). Fourth in the int-llm series.

Python ctypes binding for LC Runtime — CPU-native, bit-exact, no-GPU inference. Companion to bracoTuxbr/local-coherence.

Bit-exact MicroGPT inference and training in pure PostScript—integer-only Q16.48 arithmetic with canonical wide and uniform-F12 MGW checkpoints.

Bounded Posits at IEEE Tensor-Core Throughput on Commodity NVIDIA Blackwell, with Bit-Exact Reproducibility

CPU-native inference runtime. Local-propagation paradigm: the active region pays the cost, not the field. Bit-exact across architectures. Validated for streaming anomaly detection and audio VAD.

Header-only C++17 scipy API mirror — based on numpcpp+Eigen3+pocketfft, bit-exact alignment

One integer network implemented three times (Python, C++ for a Cortex-M7, SystemVerilog for an FPGA), machine-checked byte-identical

Float accumulation order alone flips RL reward verdicts and sampler/trainer probabilities — reproduce it on real GPT-2, then remove it with an order-independent reduction. numpy-only, runs in seconds.

Bit-exact MicroGPT inference in pure SQLite SQL—not text-to-SQL. Integer-only Q16.48 transformer math and deterministic sampling, with no UDFs or extensions.