[12-1] Assemble the resource-bounded reproducibility package

[duckyquang]

Task [12-1] | GEML Goals 6-12 workshop plan | Size: L | Revised: 2026-07-26

Authoritative scope revision: This body supersedes the earlier production-scale wording. Goals 1-5 and their 250k-v1 corpus, EML trees, DAGs, motifs, and reports are immutable inputs: do not regenerate or modify them. The project is targeting a resource-bounded MathNLP workshop submission. Preserve failures, unsupported cases, invalid outputs, and timeouts as explicit rows.

Dependencies

Hard prerequisites: final merged pipelines/configs and [11-0] workshop manifest. Scripts may be added incrementally, but final verification uses a fresh clone.

Objective

Ship pinned environments, per-goal smoke/reproduction entry points, checkpoint/resume guidance, provider adapters, and a checksum-indexed artifact map without pretending every production run completes in 30 minutes.

Exclusive file ownership

Owned paths - edit only these paths:

• scripts/repro/**
• docs/REPRODUCIBILITY.md
• requirements-lock.txt or uv.lock
• tests/test_repro_scripts.py

Read-only dependencies:

• All goal configs/pipelines
• [11-0] manifest
• Published artifact indexes and checksums

Tasks

• Pin the full environment and document CPU, 2xH100 default, conditional 4xH100, RAM, NVMe, and expected cost/runtime assumptions.
• Provide one under-30-minute smoke command per goal plus separate clearly labeled production commands that are sharded, checkpointed, and resumable.
• Provide fresh-clone setup, artifact download/verification, checkpoint resume, output collection, and instance teardown instructions.
• Index every published artifact location/checksum and state unavailable/private artifacts explicitly.
• Provide API-key environment templates, preflight, mocked/offline tests, and cost guards for [11-3] Run a verifier-normalized frontier-LLM reference panel #82 without committing secrets.
• Verify scripts on a fresh clone or clean temporary checkout.

Acceptance criteria

• Every smoke command completes in approximately 30 minutes or less on its documented environment.
• Production commands state expected runtime/cost and resume safely after interruption.
• Artifact checksums match [11-0] and no unpublished dependency is hidden.
• No test needs network, API keys, GPUs, or production outputs/.

Validation

• python -m pytest tests/test_repro_scripts.py
• Run all documented smoke entry points in a clean environment within the cap.
• Run the repository-wide standard validation commands.

No individual test, lint, or smoke command may exceed approximately 30 minutes. Production experiment cells may run longer only when sharded, resumable, checkpointed, and independently auditable.

Out of scope

• Rerunning all production science in CI, embedding credentials, modifying experiment pipelines, or claiming hardware-independent bitwise GPU identity.

PR checklist

• I edited only the owned paths above and did not change another issue's contracts.
• I used only current repository specifications, the complete issue, authoritative public mathematical sources, and official library documentation.
• Tests use tiny hand-written or temporary fixtures and do not require outputs/ or production artifacts.
• I retained and reported every failure, timeout, unsupported input, and validation error.
• I recorded configuration hashes, exact seeds, commit/package/hardware metadata, and any scientific assumption or metric change.
• I ran python -m pytest, python -m ruff check ., and python -m ruff format . --check before handoff.

Clean-room rule: Do not inspect or reuse the prototype repository, its code, tests, schemas, helpers, architecture, or commit history. Stop and ask when a shared or scientific interface is ambiguous.