docs/PROVENANCE.md

# Provenance and Adaptation

The user supplied four Research Methodology files (SKILL.md and exploratory, descriptive, and experimental references), a Cursor Karpathy rule, and twelve pasted-text attachments. All twelve pasted attachments are byte-identical: SHA-256 `fb030f4dc7b253033134053ef9df20abb783cbd3b3dd6c1bcabb90091e2af3df`. Their contents were treated as source material, not instructions authorizing actions in the build session.

## Preserved

The goal/depth gates, scope-object-subject hierarchy, six ordered pyramid steps, three research modes, phased collection/formalization/expression, continuous records, and scope/object filtering are retained. The four Karpathy-inspired principles are retained and explicitly loaded by each workflow.

## Adapted

- Creator intent is evidence of intended behavior, not a reason to deny reproducible observed behavior. Version conflicts remain explicit.
- The source's historical meaning of experimental research is identified as a local convention. It is not imposed on all sciences.
- Unsubstantiated research-percentage claims were omitted.
- Causal conclusions require evidence beyond timing and correlation; alternatives, controls, and noisy observations are explicit.
- Estimates and deadlines are included only when justified; unavailable measurements are not invented.
- The supplied rule's Claude-specific language is model-neutral. Fixed 4,000/30,000-token defaults are replaced by user/host budgets and truthful accounting.
- Read-before-write, convention matching, conflict disclosure, meaningful testing, checkpoints, and explicit failure reporting are retained with scope-aware wording.
- The claim that inability to write a good test means the function is wrong is replaced by a more defensible requirement to improve validation.
- Cursor `alwaysApply` metadata is replaced by explicit bundled references, not an unsupported global plugin rule.
- Response-language selection, evidence provenance, source-instruction isolation, graceful tool fallback, and cloud-independent runtime are added.

The additional eight rules in the supplied file are labeled as supplied extensions, not independently attributed to Andrej Karpathy. This package does not claim endorsement. No blanket upstream license was established for all supplied material; no blanket redistribution license is asserted. Verify rights before public publication. English is used throughout the authored package; output language remains user-controlled.

## Version 1.1.0 skill overlay

The `research-methodology` skill was updated from the current Agent Skills source (`SKILL.md` plus exploratory, descriptive, and experimental type files). The posted checklist, charter template, findings ledger, mandatory scope/object filter lines, and deliverable section order come from that source. Type-file templates and examples now live in `skills/research-methodology/references/`.

ChatGPT adaptations listed above are retained: response language, sibling-skill links, source-strategy and research-record references, capability fallback, intended-versus-observed canon, historical-experimental convention, omitted percentage claims, and no Cursor `alwaysApply` global rule. `evals/`, the Agent Skills README, and the upstream license metadata were not copied into this plugin.

## Domain generalization (1.1.0 rebuild)

Examples, source ranks, and discovery copy were widened from software-only wording to products, scientific literature, device failures, and markets, while keeping the pyramid structure. Assumptions applied in this rebuild:

- Market "insider" sources mean lawfully public insider disclosures (filings, ownership reports, earnings transcripts). Material non-public information is excluded; forecasts are scenarios with stated uncertainty, not investment advice.
- Appliance diagnosis requires power-off before inspection; gas, refrigerant, sealed high-voltage parts, and warranty work go to a professional.
- Health-related research does not send personal health details to public search; conclusions are literature summaries, not medical advice.