Supported scope and qualification

BrainlessLab’s active implementation is Python 3.12 and Quadrants 1.3.3. The package lives under python/src/brainlesslab/. Version history locates the earlier source and its evidence boundary.

Numerical and architecture boundaries

CPU64 uses strict ordered float64 arithmetic and is the reference policy. Metal requires explicit float32 numerics and separate backend qualification. CPU32 equation checks do not qualify Metal32.

The CI matrix requests CPU checks on Ubuntu 24.04 x64, Windows 2022 x64 and macOS 14 arm64. These are configured architecture gates. Actual CI receipts establish what passed; configuration alone does not qualify those platforms. No CUDA, Vulkan, Linux arm64 or Windows arm64 qualification is claimed.

Recorded local checks

The 7 October 2026 qualification receipts describe one macOS arm64 host with Python 3.12.13 and Quadrants 1.3.3. They are bounded software checks, not qualification of every CPU architecture or study evidence.

Receipt	What it establishes
Software checks	The recorded CPU64 test suite, lint/type checks and installed-wheel check
Backend conformance	637 short CPU32/Metal32 snapshots with declared tolerances; not long-run bit equality
Development execution	Completed local plans and bounded decoding checks; not confirmed task performance

These receipts retain their original tested source and counts. A changed release needs fresh checks. Configured CI on another architecture remains pending until its actual run passes.

uv sync --locked --group dev
uv run pytest
uv run ruff check python
uv run pyright
uv build

Keep first-call compilation, warm execution, host/device copies and recording costs separate when profiling. Record backend, dtype, thread count and cache state. Use the performance guide to plan those measurements.

Active scope

Falandays and SORN have separate equation and conformance boundaries. Historical Falandays validation does not transfer automatically to Python. SORN software checks do not establish authors-code trajectory parity or biological validity.

The active tasks include three physical control tasks, delayed cue and six additional capacity-probe families. The library has 64 difficulty cells across delayed cue and those six families. Availability is experimental software readiness, not confirmed cognition or learning.

Dense model storage is the baseline. Batching and recording are bounded. A sparse mask does not make the runtime a sparse implementation.

Workspaces, reports and sharing

Experiment workspaces connect ordinary plans, original execution records, analyses, authored interpretation and follow-up questions. Status provides bounded context. Checks verify declared references and artefact hashes; they do not execute a scientific study or accept a result.

Static reports select recorded statistics and separately executed analyses. Offline HTML includes its figures, tables and styling. Publication revisions retain declared replication material, complete renderer source and optional original run-data supplements within a 25 MiB total cap. Sharing uses the existing website and human contribution review.

The runner captures v3 source, plans, settings and declared external inputs before execution. Historical v1/v2 records retain the provenance originally recorded. Numerical reproduction comparison requires intact complete Python v2/v3 records and a separate completed execution. Its receipt covers trial outcomes, summaries and recorded contrasts under declared tolerances; it does not verify every internal state.

There is no job scheduler, automatic checkpoint recovery, arbitrary executable report plugin, notebook execution, external archive service or automatic scientific acceptance. Report checks cannot prove arbitrary analysis code has no undeclared dependencies. Omitted original data is unavailable for independent reanalysis unless it is retained in another declared part. Numerical comparison does not establish scientific acceptance. A report’s rerun flag requires matching supported receipts for every selected originating record, with both full records retained as supplements. It does not establish reanalysis or full-trajectory agreement.

Deliberate exclusions

General body composition, heterogeneous ecological populations, broad node registries, evolution, general autodiff, durable checkpoint/resume and distributed execution are not implemented in this release. Historical implementation capabilities remain available only at their original source revision.

The website plays generated Quadrants development recordings. It has no TypeScript neural or world equations, live parameter tuning or claim of learning from a selected case.

Accepted research records and historical benchmark tables retain their own provenance. The rewrite neither reruns nor promotes them. Continue with implementation exercises and research.