You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Replaced the coding evaluator's allow-default macOS Seatbelt profile with a deny-default policy.
Generated programs can read only the sealed task inputs, their output directory, explicit
runtime/toolchain roots, and Apple system-runtime paths; network access remains denied.
Moved compiler and hidden-test output capture into parent-owned pipes. Generated code can no
longer replace scorer log paths with symbolic links and induce the privileged scorer to read a
hidden test or unrelated host file. One cumulative byte limit now covers all compile and test
output for a task.
Restricted formal evaluation to Apple's SIP-protected /usr/bin/sandbox-exec, rejects control
characters in rendered Seatbelt paths, resolves Rust and Go shims to their reported native
toolchain roots, and requires post-test completion evidence before awarding a pass.
Added SECURITY.md with the
untrusted-checkpoint/generated-code threat model.
Seatbelt remains a deprecated same-kernel boundary; use a dedicated account or disposable host
for adversarial coding evaluation.
Added
Added the Qwen3.8-27B dense language-path conversion profile and AXQ 4/6-bit pack workflows.
The 4/6-bit checkpoints are Tier 1 certified; MTP variants also carry scoped Tier 2 certificates
for AX Engine 6.16.1 under the exact Qwen3.8 profile.
Added Qwen3-VL 30B-A3B Instruct MoE conversion through MLX-VLM with a protected vision tower.
Its AXQ 4/6-bit packs are checkpoint Tier 1 certified.
Added Holo3-35B-A3B grafted-MTP preparation plus an offline/online alignment ladder,
teacher-force measurement, self-distillation data preparation, and staged adaptation tooling.
Holo3 AXQ 4/6-bit direct-decode packs are Tier 1 certified; MTP Tier 2 remains closed.
Added a generated full certificate index, Qwen3.8 and Holo3 operator runbooks/evidence, and a
72-hour AX Engine 6.15.0 endurance report for Qwen 3.6 27B AXQ 6-bit.
Changed
Public certification tables now list dual Tier 1+2 products first and separate listed Hub packs
from unlisted no-MTP/evaluation records.
MLX-LM model loading in the MTP alignment path accepts current load-result tuples without
assuming exactly two return values.
GitHub release tag/version and curated-note gates now raise explicit failures and remain active
under optimized Python execution.
Updated package metadata to describe AXQuant as MLX-native mixed-precision PTQ for language and
multimodal checkpoints on Apple Silicon.
Fixed
Fixed candidate-controlled scorer log paths that could disclose files readable by the parent
process and forge coding-task completion evidence.
Fixed Rust/Go scoring when the discovered executable is a version-manager shim, while avoiding
broad read access to credential-bearing shim directories.
Preserved both lexical and resolved executable paths in Seatbelt profiles across Python 3.11+
so symlinked toolchain entrypoints remain runnable without broadening process execution.
Allowed the selected Python framework version root so GitHub Actions and other framework builds
can load their sibling Python dynamic library inside the deny-default sandbox.
Fixed strict typing and current MLX-LM compatibility across MTP dataset, teacher-force, and
adaptation commands; repaired source-preparation path normalization and grafted sidecar key
enumeration.
Prevented local runtime evidence such as runtime_check.json from being included in PyPI source
distributions.
Cleared repository-wide Ruff, formatting, and strict-mypy failures so release CI exercises the
complete current source tree.
Operator action
Coding-suite manifests bind suite version 2026.08.14.2 and the Seatbelt renderer/policy
digest. Regenerate and re-run verify-coding-suite for any suite created by an earlier release.
Existing public artifact schema_version contracts are unchanged; no plan, manifest, or
certificate schema migration is required.
Install from PyPI with python -m pip install 'axquant[mlx]==1.7.0'. GitHub release wheels, SHA256SUMS.txt, and build-provenance attestations are produced by the tag workflow.