CAMS Protocol R (CAMS-R-2026-001): preregistration, single-run results, and rubric erratum final

Abstract

Protocol F (CAMS-F-2026-001): Design Rationale Companion to preregistration_F_draft.json — 2026-09-20 Status: DRAFT. The forecast file does not exist yet; no outcome-year scoring has begun. Protocols R and R2 answered does the relational pattern generalise? — twice, under lock and key. Protocol F asks the harder question the framework has been circling since its first preregistration: does a CAMS snapshot tell you anything about the future that you wouldn't already know by assuming nothing changes? Why this design is honest The forecasts will be public before the future exists. The forecast file — every society, node, metric and horizon, computed by a frozen formula — is sealed, hashed and published at registration. There is no possibility of fitting to outcomes: the outcomes haven't been scored, and the 2030 primary outcome doesn't exist at all. This is the strongest prospective posture available to a slow-moving societal model, and it is deliberately irreversible. The claim is deliberately weak. Persistence-plus-damped-drift (origin value plus half society drift, half roster drift, clipped to the scale) is close to the weakest non-trivial forecast a framework can make. That is a feature. If CAMS cannot beat a random walk, the framework's forecasting ambition should return to development — and the FAIL verdict says exactly that, pre-written. If it can, the door opens to stronger models in later registered studies. The baselines are the sceptic's baselines. Persistence ("nothing changes") and roster-mean drift ("everything follows the crowd") are the two null forecasts any sceptic would propose. The primary endpoint is the MAE ratio against persistence, gated at ≤ 0.90 — a 10% skill margin — with an exact permutation null on the pooled skill score. The backcast is labelled non-confirmatory. Origins 2000/2010 against known 2010/2020 outcomes will be computed and published with the preregistration, for transparency about whether the model form has any skill at all in-sample-of-history. It cannot count toward the verdict; those outcomes were visible during framework development. It exists so that a FAIL cannot be spun, in either direction. What happens when At registration (now): forecast generator frozen and hashed; forecast file generated from panels through 2025, sealed, published; society eligibility list frozen mechanically (panel complete through 2025). 2028: the 2027 outcome vintage is scored (blind, R2 rubric, flat-scan), sealed, and the k=1 interim is reported — report-only. 2031: the 2030 vintage lands; the primary verdict is computed and reported in the pre-registered language. 2036: the 2035 vintage; final report. This study will outlive any single session, and it is designed to: the custodianship terms, hashes and verdict language are written so that any future agent — or any human — can execute the outcome steps mechanically and the registration will still mean exactly what it means today. Immediate next steps Freeze the forecast generator (small script; the formula above, nothing else). Freeze society eligibility (panel-complete-through-2025 rule against the verified pool). Generate the forecast file; custodian seals and hashes it. Publish the preregistration + forecast file (Zenodo v5). Wait for 2027 — and score it blind.