The diff adds only two content files: an appended history entry (append-only, stamped with date, actor, aiAssisted: true) and a new pre-registration file with zero rows, zero findings. No AGENTS.md or check-weakening edits appear, and nothing is presented as ratified — the study carries runId, model, promptVersion, and an explicit `humanReviewed: false`, satisfying §3.15 and the Phase-1 labeling rules. No citation identifiers are added (verification tooling confirms none), and the listed known candidates (Hameroff & Penrose 1996 Math. Comput. Simulation 40; Tegmark 2000 PRE 61; Hagan/Hameroff/Tuszynski 2002 PRE 65; McKemmish et al. 2009 PRE 80; Hameroff & Penrose 2014 and Reimers et al. 2014 Phys. Life Rev. 11; Craddock et al. 2017 Sci. Rep. 7; Sahu et al. 2013 Biosens. Bioelectron. 47; Derakhshani et al. 2022 Phys. Life Rev. 42) match real, well-known works with venues consistent with their titles — no internal inconsistency or invented locator. The design actively serves the constitution: freeze-before-audit (§3.12), stated-as taken from the paper's own wording rather than inferred (§3.9), symmetric grading of proponent and critique rows, exclusion of secondary restatements as independent entries (§3.10), documented silences recorded as findings (§3.11), and plain-language consistency judgment with no invented precision (§3.13). Locator quality and source support remain to be judged when the collection PR lands; nothing in this freeze is a violation.

The change introduces a pre-registered study file and a matching history entry, with zero data rows, consistent with the project’s two-PR freeze/collect discipline and §3.12’s freeze-before-audit requirement. The study design explicitly separates observations from interpretation (exact quoted numbers vs later consistency assessment) and uses one row per (publication, parameter), aligning with atomicity and explicit evidence direction once rows are added (§3.1, §3.2, §3.5). It commits to primary documents (named journal articles, book, arXiv/DOI retrieval) and to exact locators for each numeric value, while forbidding inference of the “stated-as” category, honoring §3.7–§3.9. The inclusion/exclusion rules and explicit handling of secondary summaries, generic-collapse bounds, and measurement-invoked rows respect independence (§3.10) and avoid double-counting. Documented silences and failed searches are treated as negative evidence (§3.11), and the design promises plain-language findings with no spurious numerical precision (§3.13). Provenance metadata (runId, model, date, promptVersion, humanReviewed flag) and the history entry ensure versioning and AI-disclosure compliance (§3.14, §3.15). No citations are fabricated (tooling found none to resolve), no confidential material is introduced, and no constitutional checks are weakened.

The PR explicitly implements §3.12 (Freeze before broad audit) by pre-registering the study criteria, candidates, and methodology with a strict hash before data collection begins. It mandates exact source wording without inference (§3.9) and commits to documenting literature silences and failed searches as negative findings (§3.11). The AI authorship is transparently stamped with the required metadata and marked as unreviewed (§3.15), and the referenced known candidate papers are real, well-known publications in the Orch OR corpus.

The diff is a two-file freeze: it appends one history entry and adds orch-s001.yaml with empty rows/findings/limitations, humanReviewed: false, and explicit AI stamps (runId, model, date, promptVersion). That is a §3.12 pre-registration, not a published assessment or a ratified judgment dressed as human review. No AGENTS.md edit, no identifiers or locators invented (citation tooling found none), no confidential material, and prior history entries are untouched. Known-candidate bibliographic names are protocol commitments, not Source/Evidence records carrying a verdict; the method text itself requires exact locators and stated-as labels from each paper’s wording only after collection. Ordinary freeze work under the existing checks.

This PR is itself the §3.12 freeze: it pre-registers ORCH-S001 with frozen inclusion/exclusion criteria, a literal search protocol, and pre-committed candidate dispositions, while rows, findings, and limitations are empty by design — criteria are judged before data exists. The protocol encodes the epistemic rules rather than weakening them: stated-as taken from the paper's own wording (§3.9), secondary restatements excluded (§3.10), failed searches and documented silences reported as findings (§3.11), no invented precision (§3.13), and a criteria hash that fails CI on post-freeze edits, which strengthens rather than weakens the checks. The twelve known candidates are all real, well-known items in the Orch OR literature (Hameroff & Penrose 1996, Tegmark 2000 PRE 61, Hagan et al. 2002 PRE 65, McKemmish et al. 2009 PRE 80, Hameroff & Penrose 2014 Phys. Life Rev. 11, Reimers et al. 2014, Craddock et al. 2017 Sci. Rep. 7, Derakhshani et al. 2022, etc.), with venues and volume numbers matching what I know; the citation tool found no identifiers to check and nothing in the diff contradicts itself. The history entry is append-only, honestly labeled aiAssisted: true with the AI actor named, and the study file declares humanReviewed: false — no unratified draft is presented as ratified and no AI record is presented as human. AGENTS.md is untouched, no confidential geo-project material appears, and the planned web search is editorial discovery under §3.15, not live AI added to the static site. Publication is proceeding through this multi-model gate, exactly as §3.15 requires.