Linux and Python 3.13 qualification

This receipt addresses issue #7: run the literal first-study tutorial and full offline conformance suite on Linux with a distinct Python 3.13 interpreter. It measures one local container environment. It does not qualify native agent execution, model quality, independent scientific review, every Linux distribution or other Python builds.

The issue was selected on 18 October 2026 (Asia/Seoul) because it was the remaining unclaimed engineering task without an implementing PR. Issues #19, #22 and #23 already had PRs #25, #26 and #27; #6 had a confirmed contributor; #8 required an owner rights decision. This work uses public main directly and has no dependency on those open PRs. It is maintainer-requested, agent-assisted qualification.

Measured environment

Item	Recorded value
Source	`174e200`, current main at execution
Userland	Debian GNU/Linux 12 (bookworm), aarch64
Kernel	Linux `6.10.14-linuxkit` in a local Docker Desktop Linux VM on an Apple Silicon Mac
Interpreter	CPython 3.13.16; `/usr/local/bin/python3` resolves to `/usr/local/bin/python3.13`
System tools	Git 2.39.5 and procps `ps`; exact versions are in the environment receipt
Container limits	Four CPUs, 4 GiB RAM, 256 PIDs; no host workspace, credentials or sockets mounted
Checkout	`/qualification/clean path with spaces/allagma`
Study	`work/my-first-study` under that checkout; its absolute path contains spaces
Acquisition	Anonymous HTTPS shallow, blob-filtered, sparse Git clone; no Allagma or third-party Python package installation

The base image was python:3.13-slim-bookworm, resolved to sha256:a1165e272e578941b84abc79e4ab38a0305cd12803a5c4247979ac7655f4d641. Image acquisition, system-package installation and Git acquisition used network access. The container was disconnected from its bridge network before the catalog, toy, audit and conformance commands. These commands then ran offline. The image included pip, but no third-party Python packages.

The environment receipt records the interpreter binary digest, OS, Git, sparse paths and exact source revision. The command receipt records each command’s working directory, UTC start, exit code, elapsed time and output digests. Durations are observations, not performance guarantees.

Commands and observed results

The three shell blocks were checked against the guide at the tested commit. The collection script recorded individual commands and directory transitions; it did not replace them with internal API calls. The first block was:

git clone --depth 1 --filter=blob:none --sparse https://github.com/alohays/allagma.git
cd allagma
git sparse-checkout set allagma adapters contracts methods recipes policies profiles \
  templates tools conformance examples/toy-study examples/first-research evals/context-retention
python3 -m allagma check

The remaining tutorial commands, followed by the full conformance command, were:

python3 -m allagma toy --destination work/my-first-study
python3 -m allagma campaign audit --study work/my-first-study --campaign toy-v1
python3 -m unittest discover -s conformance -v

All Python commands used the same recorded Python 3.13 executable, including the child processes launched by the toy and tests. The results receipt binds the catalog, complete toy, repeated audit, conformance log and evidence inspection. No expected scientific hash or test assertion was changed to obtain the result.

Check	Exit	Observed result with `procps` present
Catalog	0	All 23 modules pass
Complete default toy	0	26 successful attempts, one deliberate failure, one interruption
Explicit repeat audit	0	Pass; 521 references, no findings or stale claims; new `review-002`
Full conformance	0	192 tests pass in 49.037 seconds; zero skips
Separate raw-data recomputation	0	24 confirmation replicates and all manuscript values/claim directions agree

The full passing conformance log is retained. The manifest binds the compact receipts, including the read-only inspection script. That script is evidence-collection support, not a new Allagma runtime command.

Missing system prerequisite and retained attempts

The first slim-image run had Git and Python but no ps. The catalog, complete toy and 521-reference audit passed; full conformance ran 192 tests and failed with two failures and 15 errors in resource-supervision and related cases. The retained failure receipt and failure excerpt identify the affected tests and missing executable. This is a missing system prerequisite, not evidence of Python 3.13 incompatibility.

The existing resource-supervision contract already requires ps, but the contributor and conformance entry points did not name it. Those entry points now document the requirement. On minimal Debian/Ubuntu, install procps before running the full suite; the needed interface is:

ps -axo pid=,ppid=,pgid=,rss=,stat=,lstart=

A second container used the same pinned base image plus Git, CA certificates and procps, with a new anonymous sparse checkout and a new study. The first container was stopped to terminate the failed run’s remaining child processes; its filesystem and study were preserved. This changes the environment to meet the existing prerequisite, not the tested Allagma source. Neither study’s 28-attempt/60-second ceiling was increased.

An earlier collection-script self-check also caught an extra + in the sparse path list before any Allagma execution. That setup attempt and its correction are recorded separately; the measured runs use the exact guide path list.

Manuscript and evidence inspection

The generated manuscript was read alongside summary.json, claims.json, the raw manifest and the seed-104 raw measurement. Its three Markdown links resolve to the figure, claim ledger and analysis record inside the study. The claim ledger’s direct references were checked against their digests, and the campaign audit followed the complete 521-reference chain and regenerated the report.

The report gives sample-mean MSE 0.01529948, biased-estimator MSE 0.08040365 and paired difference 0.06510417, with approximate interval [0.03985104, 0.09035729]. It keeps the 24 independent confirmation seeds distinct from retries and the two pilots. The average-effect claim is supported; the universal claim is contradicted by seeds 104, 108, 118 and 121. Seed 104’s mean is −0.34375, so adding 0.25 reduces its squared error, consistent with that counterexample. The report calls the interval a normal approximation and limits the result to this synthetic known-answer example.

As an additional check, the existing tools/audit_toy.py recomputed the raw sample arithmetic, CSVs, figure points, claim directions and report numbers without importing the Allagma or study helpers. Its rational-arithmetic receipt is deterministic verification of this example, not scientific or model-quality qualification.

The repeat audit appended review-002 and an event, updated the two documented mutable state/audit pointers, and preserved the other 753 retained files by SHA-256 and nanosecond mtime. This inventory excludes interpreter bytecode caches, as shown in the retained inspection script. Original attempts, bundles, analysis and review-001 remain intact. The study finished at phase audit, execution status completed, assurance deterministic. Only compact sanitized receipts are committed; the complete studies remain in the local stopped containers.

Main advanced during qualification

PRs #26 and #27 merged while this work was running. The initial receipts above remain bound to 174e200 and are unchanged. A fresh sparse checkout then tested da9ea458, the new main containing both fixes, in the same Linux/Python 3.13 environment.

All tutorial commands passed again: 23 catalog modules, the complete 26/1/1 toy, and the 521-reference audit. All 206 conformance tests passed, without skips, in 51.861 seconds. The new study’s manuscript values, links, rational recomputation and preservation checks also pass. Its study and receipt are separate from the original run. The integration receipt records the source-selection guard, exact commands and outputs; the full conformance log is retained.

PR #25 then merged. A further fresh sparse checkout at f7e32a7 also passes all tutorial commands and 217 conformance tests, with zero skips, in 55.100 seconds. This includes the merged toy-paper and audit-binding regressions. The follow-up receipt and full log preserve that run separately. The toy again retains 26/1/1 attempts, a passing 521-reference audit, unchanged original evidence and the same reported values.

The documentation adds a discoverable evidence page and links from the tutorial and host-support guide. The production build, internal links and all 18 existing desktop/mobile browser checks pass. The new page was also inspected in light and dark themes at 1440- and 390-pixel widths; its links work and the page does not overflow the viewport. These are documentation checks on the PR, not additional platform or native-host qualification.

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