Skip to content
Allagma

Release 0.3.0rc2

The public-launch package adds a verified offline tutorial, a Starlight documentation site, a real-software demonstration film, compact source distribution and contributor/support materials. See owner actions for current publication decisions; the repository and documentation site are now public. This document retains the candidate’s original qualification evidence. Later main-branch features and current contributor work are listed in the triage record. These presentation changes do not rewrite the frozen comparison or promote the candidate to a stable release.

The three required study packages, twelve-session comparison, clean-checkout reproduction and final source acceptance have passed their documented gates. All twelve sessions completed the required science; ten delivered complete evidence packages. The comparison explains the two review-binding gaps and the limits of this small experiment. This release is a source candidate, without a package-index publication or general model-quality claim. The acceptance ledger and publication receipt record the verified scope and successful source publication to origin/main.

Reusable workflow and corrections

Section titled “Reusable workflow and corrections”

The brief/material/resource interface prepares isolated study-owned workspaces and invokes fresh Codex sessions through pinned methods and adapters. Local resource supervision retains failed/interrupted attempts, separate setup and compute accounting, source snapshots and public native usage. Scientific code, metrics and hypotheses remain study-owned. Default conformance and the toy study remain offline and standard-library-only.

The frozen rc1 broker supplied MPS high watermark 0.2 but left PyTorch’s low watermark at an incompatible default. The new broker explicitly supplies low watermark 0.1 while retaining the high cap and disabled implicit fallback. The actual full-broker GPU check passes known forward/gradient values, AdamW, checkpoint/resume and real filesystem/network denials without worker-side environment repair.

A fast worker could finish after a storage sample and be accepted before its final footprint was measured. Both supervisors now record final persistent storage and reject nominal success above the cap. Existing failure reasons are preserved, with the storage violation also recorded. The same before/after process fixtures reproduce the old problem and detect the corrected outcome. The native-adapter fixture uses a fake CLI and does not qualify a hosted model.

The current evaluation scorer accepts the specified inline curve list and the original path representation. The frozen scorer and its original failures remain retained. Postprocessing preserves terminal queue evidence and internal aliases, recognizes named manifest entry lists and streams large archive restoration. These changes preserve original candidate outcomes and scientific artifacts.

Concrete release evidence

Section titled “Concrete release evidence”
  • I1–I5 acceptance: all five scenarios and 113 offline standard-library tests pass without skips. The complete toy retains 26 successes, one failure and one interruption; its transitive audit checks 521 evidence references. The 5,944-file archive was restored and its pinned toy helper reanalyzed the results successfully.
  • Independent toy calculation: exact rational arithmetic verifies primary/sensitivity data, claims and manuscript.
  • Three illustrative packages: CORE r01, modular r04 and EMA r07, selected prospectively as the first verified Allagma run per task. Their sources, protocols, attempts, analysis, reports and reviews are retained.
  • Full clean-checkout CULP reproduction: new environment, all three original scripts, independent graph checks and a separate raw-data recomputation pass. That checkout is the frozen workflow before rc2 repairs; final repairs have their own current-source regressions.
  • Historical freeze/archive audit and actual cohort controls pass for all twelve runs. Five optional scientific scorer tests, six archive transport tests and four comparison tests also pass.

The accepted distributable source is sha256:a7a67e09548e40442f2ff72ba63898f66e02368046132a183588ebebdbc1a196. The receipt separately captures tests, examples and contributor materials.

Migration and evidence boundaries

Section titled “Migration and evidence boundaries”

Record contracts remain version 0.2 and the scaffold remains version 1. Old bundles/campaigns retain their exact helpers; a new release does not rewrite an old campaign or remove its failures. Use Check/Plan/Reconcile/Validate/Adopt at a campaign boundary, or prepare a new research workspace. Paired seed handling and study-owned minimum confirmation counts introduced in rc1 are unchanged.

Native model configuration remains temporary. No project-level model or review model override is introduced, and user GUI/personal model choices are preserved. Scientific computation stays local; model usage uses the existing account.

The comparison estimates the frozen rc1 conditions, including the shared MPS defect and each candidate’s autonomous fallback/repair. It does not estimate the performance of the corrected GPU path. Original outcomes, later controller corrections and final release checks remain separately identified.

The real native evidence is scoped to CLI 0.160.1, inherited gpt-6-astra / max settings and this macOS M4 Pro environment. Claude Code remains contract-tested, not natively qualified. RSS excludes some GPU allocations; polling and final logical-byte checks do not provide instantaneous quotas or adversarial isolation. Scientific critique is explicitly scoped author/controller review, not independent peer review. No experiment was required to produce a favorable scientific finding or an Allagma advantage to count as complete.

Canonical source · Generated from the maintained repository document at this build’s commit.