How IESET is produced

IESET is a human-directed research framework. Large language model agents draft specs, run scripts, and write documentation at high tempo. Models are workers, not authorities. The evidence substrate — committed hypothesis YAMLs, publisher manifests, run artifacts, and git history — is the source of truth, not model prose.

Pipeline

1. Brief — a human or control-plane brief assigns work (new hypothesis, data vintage, rerun, audit).
2. Agent draft — models (API and/or local) write YAML specs, runner scripts, or docs against the repo.
3. Pre-register — the falsification rule is committed before (or with explicit ordering against) the run artifact; CI checks ordering.
4. Run — estimators execute on pinned public-data vintages; diagnostics and result cards are written under engine/runs/.
5. Gate — schema validation, pre-registration checks, and (where configured) estimator floors prevent broken arithmetic from counting as directional wins.
6. Publish — the static site build renders cards from the committed corpus.

What this means for readers

• Many specs are authored and graded at machine tempo (seconds between pre-registration and run). Git ordering still holds; human deliberation between rule and result may be minimal — the site says so when timestamps are close.
• Corpus composition has reflected production targets (verdict census goals) as well as research importance. Prefer the Featured tier and high-integrity scoreboard weights over raw win counts.
• Verdicts are provisional research artifacts, not peer-reviewed findings. Quote them with sample size, identification quality, and tier.

Where to look

• Methodology invariants
• Author disclosure
• Scoreboard (raw vs attribution gates)
• Public repository
• Control-plane notes in-repo: engine/brain/README.md

This page exists so nobody has to clone the repository to discover that the corpus is agent-assisted. Concealment would be worse than disclosure.