START is a research and curriculum tool that assembles configured research inputs, provider-assisted content, translations, and derived visualizations into reviewable filesystem artifacts. Its central design claim is operational: an artifact is safer to publish when its identity, inputs, evidence status, quality result, and generation history are visible at the point of review. START therefore uses a typed, resumable pipeline with atomic writes and a run manifest rather than treating a successful process exit as evidence of a valid publication.
This document is a manuscript-facing engineering note. It does not claim that generated content is independently verified, that synthetic foundation material is live research, or that the pipeline has demonstrated learning outcomes in human subjects.
The contribution is a compact operating contract for a single-user research and curriculum workflow:
acquire → prepare → process → parse → render);The system is not a multi-user platform, a fact-checking service, an automated ethics decision-maker, or a substitute for domain-expert review.
The methods protocol is recorded in Methods and Evidence Protocol. The first-principles function is to make reviewable artifacts, not to maximize the number of generated files. The scientific method is applied as a sequence of falsifiable engineering hypotheses, controlled offline fixtures, measured quality outputs, and iterative repairs.
The primary unit of analysis is an artifact record. It contains a stable item identity, producing stage, regular-file path, content hash, status, and provenance. The primary unit of execution is a run manifest. It records the configuration digest, stage results, provider and prompt metadata, usage, quality status, errors, and artifact records.
START makes a deliberate distinction between:
Synthetic_FEP-ActInf.md is foundational synthetic material. It is useful for
offline execution but is not a live source and is not independently verified.
The same distinction is carried into generated metadata and publication-mode
validation.
Results must be generated from the current checkout rather than copied into a static claim. Use:
uv run start-gate full # suite + branch-aware coverage
uv run start-gate docs # strict docs build + repository/link validation
uv run start-gate artifacts # provenance, output contracts, offline fixture replay
uv run start-gate numbers # generated timings/coverage page is current
The current arm wall times, test selection counts, and coverage total are in
_generated/gate-timings.md, written by
uv run start-gate measure. They are intentionally not copied here because
counts, dependencies, and generated inputs change.
The present record evaluates software contracts and artifact integrity, not learner outcomes or provider factuality. A future study should preregister a human-review rubric, define a comparison condition, measure reviewer agreement, and obtain appropriate consent before evaluating educational effectiveness. Any live pilot must be explicitly budgeted, source-reviewed, and kept separate from the offline reproducibility evidence.