Skip to content

ai-usage.md

Latest commit

 

History

History
79 lines (63 loc) · 3.67 KB

ai-usage.md

File metadata and controls

79 lines (63 loc) · 3.67 KB
Raw

AI Usage

This statement covers the software artifact in this repository: the locked instrument, the scripts, the tests, the provenance wiring, and this documentation. The accompanying paper (greenrisk_paper.pdf) carries its own AI usage statement covering the manuscript.

It is published here for the same reason every score in this project ships with an activation trace and a provenance graph: a result should carry a record of what produced it.

Where generative AI was used

Generative-AI tools were used as an execution and review layer, under author direction:

  • Agent-based coding assistants supported implementation, debugging, testing, and the reproducibility engineering of the pipeline — the artifact producers under scripts/, the smoke and property tests under tests/, the PROV-O serialization, and the consolidation of internal working notes into the public documentation under docs/.
  • Language models accessed through APIs assisted with organizing and analyzing a large body of candidate references, and with reviewing prose for clarity, structure, and internal consistency.

All AI-assisted code, analyses, and text were reviewed by the authors before inclusion. All cited sources were checked against the original publications.

Where it was not used

The following were determined by the authors and are not AI-generated judgments:

  • the research question, the domain analysis, and the selection of the four constructs;
  • the signal mappings, the membership functions, and the 17-rule Mamdani base, including the DL-001 correction that reshaped the spine;
  • the evaluation protocol — the decision to freeze the instrument, the blind pre-registration of the held-out contrast set, and the one-shot scoring discipline;
  • the interpretation of the evidence, the scope limitation the results expose, and the conclusions drawn from them.

Every one of these decisions is recorded in decisions.md with the evidence that produced it and the date it was ratified. The AI systems did not act as autonomous researchers or authors, and were not treated as a source of evidence.

Why this is checkable rather than merely asserted

The project's existing discipline is what makes the division above verifiable rather than a claim about intent:

  • every instrument decision was logged before it was relied on, and the alternatives that were rejected are logged with it;
  • the instrument was frozen at rulebase-locked-v1 (commit a40288a) before the held-out case set was inspected, so no post-hoc adjustment — by a human or a model — could have fitted it to the test;
  • the held-out set was scored once, against the pre-registered stratification published in Appendix A of the decision record;
  • the five ClimateBERT classifiers are pinned by revision, and every run emits a W3C PROV-O graph binding its results to those revisions and to the instrument tag;
  • the claims this documentation makes about the rule base are asserted in tests/property_test_rule_base.py, so a change that breaks them fails rather than passing silently.

AGENTS.md exists for the same reason: it constrains what an agent contributing to this repository may treat as a routine change, and states explicitly that any edit to MODEL_REGISTRY, SIGNAL_MAP, the membership functions, or RULES is an instrument change subject to the logging and re-validation discipline above.

Responsibility

The authors retain full responsibility for the software, the reported results, and the citations in this repository.