Layev is an open research implementation combining Kev's learned option scorer and shared-state question execution with Laya-inspired calibration and reward training. The Python package and CLI retain the kev-laya / kev_laya names. The code provides typed Choice, Score, and Noul decisions, bounded parallel question batches, training, evaluation, calibration, telemetry, and an Overwatch integration payload.
Status: The source has passed CPU fixture tests in its delivery environment. It is a research candidate, not a validated Jev replacement. Native Qwen/tokenizer and CUDA checks, trained 32k/64k context, general predictive quality, Jev-relative quality/cost/latency, and a complete live Overwatch pipeline remain open. See the next-stage implementation prompt for the acceptance plan.
The handoff recorded 280 passed, 5 skipped on Linux CPU. A fresh Windows run here also passed 280 tests with 5 skips, including the real Rust byte-level tokenizer test after correcting its trainer input type. These are separate from the historical archive result. The strict 24-configuration audit, exact interrupted/resumed continuation, real 16-row question batching, fresh calibration/export and installed-wheel loopback checks have been exercised. The guarded integration upgrade, source/index checks and group-aware uncertainty reporting are documented in release readiness. See current validation and environment boundaries.
These results do not close pinned Qwen/CUDA, trained 32k/64k context, representative predictive quality, Jev comparison, or the full Overwatch backend/report/UI pipeline. Remote CI and merge must be checked separately from local validation. The repository's existing lock is retained unchanged.
src/kev_laya/: model, training, serving, telemetry, and checkpoint code.tests/: software and numerical regression tests.integrations/overwatch/: guarded adapter payload and instructions for the existing Overwatch application.docs/: architecture, telemetry, correctness, and validation notes. The archive README (link-normalized copy) is retained for provenance; its relative evidence paths refer to the separate delivery ZIP.licenses/andNOTICE: upstream attribution and license texts.
The source came from Kev_Laya_Correctness_Integration.zip (delivery version 0.1.0+stage3). That ZIP contains the larger historical evidence and trained fixture artifacts. Those artifacts and any private environment values are not part of this source repository. The original 45-feature research matrix is preserved byte-for-byte with its Layev evidence mapping. The mapping records separate source, test, native, quality, and Jev evidence states. It does not establish Jev parity.
Use Python 3.12 or 3.13 in an environment separate from Overwatch. requirements-cpu.lock records the delivery's Linux CPU closure; it is not a validated Windows or CUDA lock.
python -m venv .venv
.\.venv\Scripts\python -m pip install -e '.[dev]'
.\.venv\Scripts\python -m pytest -q
.\.venv\Scripts\python -m compileall -q src tests scriptsOn POSIX, use .venv/bin/python in place of .venv\Scripts\python. Native requirements and prerequisites are described in stage-three correctness. Do not run native training or cloud jobs without checking the required hardware, licensed data, and cost boundary.
The Overwatch integration guide documents its separate supported Python and private dependency environment. The current Windows payload uses ModelRunsView.tsx to avoid a case-insensitive filename collision with modelRuns.ts.
Apache-2.0. See NOTICE and the retained Kev and Laya licenses.