ByteLite LLC · Research and development · Helena, Montana, USA
We build computing you can check: data that rebuilds byte for byte, and machine cognition designed to answer only from knowledge it can trace.
Evidence to date is internal. Nothing has been independently verified. No product is on sale, and checkout is switched off. How we label evidence
Tested in our labs
Measured results from our own automated test runs. Each figure says exactly what it covers.
4,515 / 4,515
byte-exact rebuilds
Each one was decoded and matched to its original byte for byte before it was accepted. Separate 8 MiB and 64 MiB files also rebuilt exactly.
Stage: tested in-house. Compression release of 12 October 2026, one machine, 301 small test inputs. Exact rebuilding is not compression: every output was larger than its input. For example, a 4,096-byte text sample became 15,541 bytes, and a separate 67,108,864-byte (64 MiB) test file became 301,984,168 bytes. Effective compression has not been demonstrated.
4,758 / 4,758
automated checks passed (Release run)
In the recorded Release run, all 103 test suites passed, including all 4,758 checks. The Debug run also passed 103 of 103 suites; its check count is not claimed here.
Stage: tested in-house. Cognition release of 15 October 2026, certified for build and automatic tests only, one machine. It is not the current development branch, which is newer and not yet certified. Tests that need large language collections were not exercised.
By design
gaps reported, not filled
When nothing is on record about a concept, the system replies that it does not know and asks for more information rather than inventing an answer. This holds in the cases our tests cover.
Stage: tested in-house, tested cases only. Cognition release of 15 October 2026. Not yet shown for every phrasing: handling of ambiguous input and the grounding of claims in replies are still being completed.
39 / 39
trace checks passed
Answers drawn from knowledge are built to carry the record they came from, its stated source and its status, and each request runs against one fixed, recorded knowledge version.
Stage: tested in-house. Cognition release of 15 October 2026, 39 trace checks. Sources are recorded as supplied. How strongly a record is labelled as established, and how a source is matched to a claim, are being tightened, so traceability is not yet shown for every answer.
16,777,216
colours matched, six field computations
The accelerated paths (SSE4.1 and AVX2) gave byte-identical results to the reference path for every possible 24-bit colour, in each of six colour-field computations.
Stage: tested in-house. Vision release of 10 October 2026, one machine. This covers six colour-field computations, not image processing in general. Synthetic and test images only.
213 / 213
architecture-level tests passed
The audio research architecture passed all 213 of its tests.
Stage: tested in-house. Audio architecture release of 10 October 2026. Architecture-level testing only: no speech, music or audio analysis yet.
Research programs
Building AI designed to answer only from knowledge it can point to, and to say "I don't know" rather than guess.
Representing data so it rebuilds exactly, byte for byte, every time, with every byte accounted for.
Describing images and sound in exact, repeatable structures that other systems can rely on.
Where it is going
Information
Preserve information exactly and describe it in structured, repeatable forms.
Understanding
Establish what is known, how concepts relate, and what is still missing.
Governance
Keep behaviour inside explicit constraints, with provenance and validation.
Applications
Products may use research results only after they are validated.
Independent evaluators, research partners and licensees: pick a result and check it under your own conditions.
