title: README
emoji: 🧾
colorFrom: gray
colorTo: blue
sdk: static
pinned: false
Anomly
Anomly builds exact, verifiable AI computation: inference, rendering and training whose arithmetic is reproducible bit for bit, with cryptographic receipts a third party can re-execute.
The core is a 256-bit exact accumulator (a quire) paired with narrow posit-style operands (b-posit). Every multiply-accumulate inside a matmul is an integer operation with one rounding at the output, so the same weights and inputs give the same bits on an x86 CPU, an NVIDIA GPU or an ARM board, and an independent implementation can reproduce a served answer or a training step from the receipt. Anomly is a member of the NVIDIA Inception program.
Measured, not asserted
- Hosted inference is not reproducible today. The same eight prompts submitted ten times to ten hosted endpoints (temperature 0, seed where accepted, provider pinned) came back different every time; the closest endpoint gave three distinct answer sets in ten. One workload, one afternoon, a property of batched float arithmetic and not a vendor fault. Runs and tooling are published in the INVAR repository.
- Cross-hardware identity. The same model and prompt produce a bit-identical whole-graph hash
on three substrates (x86, an NVIDIA RTX 5090 and an ARM Cortex-A53 board), covering every traced
row of every layer; a Tenstorrent Blackhole card and a Xilinx Alveo U200 FPGA independently
re-executed the challenged rows and accepted every one. The identity board with byte counts ships
in the
identity/folder of the everywhere-model repositories below. - Further substrates. The exact profile runs the whole graph on an Apple M4 Pro GPU (Metal) with dumps identical to the x86 build; an IP-safe port of the exact-quire kernel is bit-exact on an AMD Instinct MI300X; a RISC-V build under QEMU reproduces the same whole-graph hash.
What is here
- INVAR (Apache-2.0, v0.1.22): a local, OpenAI-compatible LLM endpoint that emits a hash-chained receipt for every completion, with attestation binding, TPM signing, SCITT-style statements, a transparency log, per-matmul spot-checks, a six-level determinism certifier for any endpoint, and two independent reference re-executors (Python and Go). Under its exact profile the whole model graph is bit-identical across CPU, CUDA and aarch64 for the llama, Qwen2, Gemma 3 and Mistral families; the arithmetic is written down as a specification with conformance fixtures.
- The everywhere model: qwen2.5-coder-1.5b-everywhere-exact-w8 and its 0.5B sibling for 2 GB boards: a coding model fine-tuned on execution-verified data, served through the exact profile, byte-identical on x86, NVIDIA and ARM, with the training run, the served answers and the evaluation scores all carrying re-executable receipts.
- qwen3.6-27b-exact-w8-seed4: a paired QLoRA of Qwen3.6-27B, one arm in bf16 and one with every linear GEMM run as exact 8/10-bit integer tensor-core products, reaching the same held-out loss across five seeds, with both training runs bound into receipt chains and a bit-exact replay segment. Parity, not lift, is the claim.
- b-posit8 GGUFs for the exact profile, one repo per model with the upstream licence on the card: SmolLM2-135M (the one INVAR's fixtures and vectors reference), SmolLM2-1.7B, Qwen2.5-0.5B, Llama-3.2-1B, Gemma-3-270M, Mistral-7B-Instruct-v0.3.
- open-bposit: the open, bit-reproducible number format behind all of the above, with a reference implementation, an exhaustive conformance suite and a verified hardware block.
What we do not claim
Exactness is not confidentiality (that is a TEE property INVAR composes with, not replaces), and reproducibility does not make a model better. The b-posit8 tensor type on its own, in a stock runtime, is not reproducible inference: it is half of a pair with the deterministic elementwise profile, and we measured that the type alone diverges across machines like any other quantisation. On NVIDIA GPUs the non-exact path is still several times faster on small models. Every number we publish carries the measurement that produced it and the limits of what it shows.
Contact: https://anomly.com/contact