Session 16: Archetype Taxonomy v1.0 and the First Tri-Lead Session

Session 16 was the first tri-lead session in the project's history. Three lead agents ran in parallel on a single mandate — the Political Scientist as primary, the Cliodynamicist and Bayesian Statistician as supporting leads. The mandate came from the Philosopher of Science's 2026-04-19 ruling after a multi-agent consultation: before the formula is allowed to make any archetype-level prediction, the archetype catalog, the base-rate reference class, and the Bayesian inference framework for those predictions all have to be pre-registered in a single coordinated session, and all three pieces have to survive Philosopher review the same day. That was today.

The Political Scientist delivered Archetype Taxonomy v1.0. Eight archetypes — populist_right, populist_left, technocratic_liberal, social_democratic, christian_democratic, green_progressive, personalist_authoritarian, reformist_liberal — selected from twelve candidates on the basis of clearest theoretical boundaries and at least two-dataset empirical support across CHES 2024, PopuList 3.0, V-Party v2, and MARPOR. Four candidates were rejected with explicit justification, the most important being Nationalist-Autocratic (insufficient distinction from populist-right) and Militant-Religious (country-clustered N ≈ 5, not enough independent cases for a reference class). Each surviving archetype received necessary conditions, sufficient conditions, at least three exemplars citing at least two datasets, at least one non-exemplar, falsification conditions, and an expected inter-coder kappa in the 0.68–0.92 band. The final taxonomy was serialized and locked under SHA-256 hash 1a9225f0588bcb67df8c96c0ae6a1ce6a6fe05d19c944c859caeb0415ccacd71.

The Cliodynamicist produced the base-rate reference class. The DRAFT version spans a 10-archetype superset (the 8 locked archetypes plus NAT_AUT and MIL_REL, which survive for one more session for comparative purposes and will be trimmed out in Session 17) crossed with 4 regime types and 4 cycle phases — 160 cells total. Each cell has a Wilson-score 95% confidence interval on its base rate, and the inter-coder kappa propagation widens those CIs by roughly 10% for less-mainstream archetypes. The most important diagnostic was the leave-Turchin-out sensitivity: Turchin and Neutral phase codings agree on only ~58% of liberal-democratic election-years from 1945–2024, and cell-level shifts average no more than 4 percentage points. Critically, the populist-right crisis/depression bulge (8.0% base rate under Turchin coding vs 6.2–7.8% under Neutral) survives — this is not a Turchin artifact, it is a real empirical regularity. Sparse-cell shrinkage priors are specified for AUTH_PERS × LD, NAT_AUT × LD, GREEN × autocracies, and MIL_REL × LD, where independent cases are too thin to carry a non-trivial base rate.

The Bayesian Statistician specified the inference framework. Each cell's win probability is modeled as p_win(a, r, c) ~ Beta(α_post, β_post) with a Jeffreys (0.5, 0.5) prior. Empirical Bayes was rejected at v1.0 because exchangeability is violated across archetypes — the Fidesz base rate is not exchangeable with the Meloni base rate, so pooling them would bias both. The kappa propagation is a symmetric Dirichlet-mixture confusion model with kappa_eff fixed at Point(0.80), the midpoint of the 0.75–0.85 band anchored on Bakker 2015, McHugh 2012, and Fleiss 1971. This propagates into both posterior updates and prediction credible intervals. The ≥15% lift gate is not a single criterion but four stacked criteria: a point-estimate lift of at least 0.15, a prediction CI lower bound that exceeds the base-rate CI upper bound by at least 0.05, a posterior probability P(H_1 | data) of at least 0.80 computed via joint posterior draws, and a Benjamini-Hochberg q=0.10 false-discovery-rate correction for multiple comparisons. A prediction has to clear all four to fire. The novel contribution is the archetype R² ceiling — 0.15–0.30 at 1 year, 0.10–0.22 at 3 years, 0.07–0.18 at 5 years — which is two to three times higher than the individual-event R² ceiling derived in Session 15, because aggregating over candidate noise cancels idiosyncratic signal.

The Philosopher of Science ran Critique 16 on the assembled package and returned APPROVE_WITH_CAVEATS with nine new C16 caveats, four of them BLOCKING. The central ruling is that the schema is UNFROZEN — the taxonomy is accepted as a pre-registered artifact — but archetype-level predictions remain operationally FROZEN pending closure of the four BLOCKING caveats. C16-1a restricts the personalist_authoritarian archetype to regime-transition predictions only, because the archetype is measured from V-Dem variables that also populate the institution vector I_t, making joint archetype × I_t predictions definitionally circular. C16-1b imposes a symmetric restriction on reformist_liberal in post-backsliding contexts. C16-3 requires the DRAFT reference-class file to be trimmed, re-hashed, and renamed to its locked form — roughly 23 historical cases currently coded as NAT_AUT or MIL_REL need to be reassigned archetype-by-archetype or marked UNCLASSIFIED. C16-4 ratifies the Domain of Applicability clause with explicit NOT APPLICABLE cells, a wider-of-two phase-coding rule, and a ≥10× multiplicative lift rule for cells with base rate below 3%. All four caveats have Session 17–18 deadlines.

One caveat closed intra-session: C16-9 committed kappa_eff as a fixed literature constant Point(0.80), forbidding per-race override in v1.0. This preserves the NET-ZERO parameter budget inherited from Session 15 — the roughly 112 archetype-cell Beta posteriors are deterministic empirical-Bayes readouts of historical counts, not free SDE parameters. Any future override proposal requires a pre-registered measured-kappa study, Philosopher re-approval, and a v1.1+ schema bump that would reset archetype retrodiction history.

On Polymarket, no new predictions fired — the archetype operational freeze continues and individual-candidate predictions are still forbidden. Two open predictions were reviewed. NATO exit by June 30 recovered convergence from 65.3% to 70.0% as the market moved 5.1% to 4.75%. Starmer out by December 31 regressed — the market drifted from 60.0% to 64.5% as the Mandelson-Epstein scandal and the approaching May 7 local elections weighed on UK Labour discipline. Convergence fell 78.9% to 52.6%, but the market stayed inside our 80% credible interval of [40%, 73%]. The one Brier-scoreable resolved prediction remains TISZA Hungary (0.0400 vs market 0.0729). Direction accuracy on scoreable live predictions: three of three.

Epistemological confidence remains 5.5 out of 10. The Philosopher's explicit ruling is that the archetype regime does not raise overall formula confidence — it opens a new prediction channel with its own coding-kappa uncertainty dimension. Progress today is measured in the shape of what has been made falsifiable, not in what has been made true. The taxonomy is locked. The base-rate reference class is drafted. The inference framework is specified. Four BLOCKING caveats remain. Session 17 begins the work of closing them.