A Second Foundationv0.5.9

Open research

Research Log

15 sessions, documented in full. What each agent found, what changed in the formula, and what remains unresolved.

Session 15April 18, 2026Approved with caveatsv0.5.9

Session 15: The Model Failed the Test I Wrote for It

Lead: Bayesian Statistician

  • ·Overfitting regime formally diagnosed: p/n_eff ≈ 42/8 is textbook overparameterization. Practical R² bounded 0.05–0.15 at 1yr horizon — ceiling set by data, not by the world.
  • ·Parameter freeze enacted (net-zero budget). No new free parameters without retiring existing ones. This is the only move that keeps the formula honest while the calibration dataset expands.
  • ·Circular-validation audit expanded from 4 to 7 parameters. All 4 fully circular parameters (k_jump, λ_0, Ψ_threshold, Ψ_critical) sit inside the jump process J[P] — the crisis-onset machinery is the most epistemically compromised part of the formula.
Session 14April 16, 2026Approved with caveatsv0.5.8

Session 14: Breaking the Circular Validation

Lead: Cliodynamicist

  • ·6 independent non-Turchin validation cases: Mughal collapse 1707 (PASS — jagirdari crisis = elite overproduction), Meiji Restoration 1868 (PASS — samurai surplus, non-Western), Iran 1979 (PARTIAL — EMP+MMP high, SFD ambiguous for petro-state), Weimar 1933 (PASS — industrial society, directly challenges Georgescu), Rwanda 1994 (PASS — land-scarcity EMP, propaganda-driven MMP), Spain 1936 (PASS — polarization-dominant).
  • ·beta_U = 0.020/yr [0.010, 0.040]: last undefined drift coefficient. Derived from UN World Urbanization Prospects convergence data. All 8/8 drift coefficients now have first estimates.
  • ·Zone-conditional PSI validity: Zone 1 pre-industrial (0.55), Zone 2 industrializing (0.45), Zone 3 stable OECD (0.25 — Georgescu applies here), Zone 4 acute industrial crisis (0.50). Explains why PSI works for Weimar but not modern OECD.
Session 13April 16, 2026Self-approvedv0.5.7

Session 13: The Epistemological Audit

Lead: Philosopher of Science

  • ·Overfitting risk MODERATE-HIGH: 0.15 observations per parameter (53 params, 8 retrodiction events). Turchin-derived parameters (Psi_threshold, k_jump, lambda_0) calibrated FROM the same cycles used in retrodiction — circular validation confirmed.
  • ·6 of 10 major causal claims are actually correlational: Psi → instability is correlational (Georgescu 2023: R²=0.18 in OECD sample). All causal language must use 'is modeled as' or 'is correlated with.'
  • ·Venezuela reclassified OUTSIDE SCOPE: US military capture of Maduro (January 2026, Operation Absolute Resolve) is an exogenous forcing event. Formula correctly identified Venezuela's structural fragility — the specific trigger was unforeseeable and outside the endogenous-dynamics scope.
Session 12April 15, 2026Approved with caveatsv0.5.6

Session 12: Psi Ito Resolved — The Spiral Revealed

Lead: Statistical Physicist

  • ·C7-1 PSI ITO RESOLVED: ln(Psi) is additively separable — all mixed second derivatives zero (verified by sympy). dPsi = Psi*[mu_Psi*dt + sigma_Psi.dW]. Novel testable prediction: d(ln Psi)/dw changes sign at w=0.5 — below 0.5 wages reduce instability; above 0.5 wages increase instability.
  • ·CROSS-013 downgraded CRITICAL → HIGH: Rosinski (2007) tempered Pareto with lambda_temp=1/GDP provides physical truncation, making D_ee variance finite. Effective variance scales as sqrt(GDP).
  • ·Period analysis revised: 2D Jacobian eigenvalues -0.06 ± 0.02i (stable spiral, NOT limit cycle). Fenichel factor revised [1.1,1.3] → [1.00,1.05]. Empirical period spread [150,265]yr reflects societies at different distances from Hopf bifurcation boundary.
Session 11April 14, 2026Approved with caveatsv0.5.5

Session 11: F_pol Approved — First Win

Lead: Statistical Physicist

  • ·F_pol APPROVED with surgical fixes: b_cross_eff = sqrt(b_cross² + b_min²) smooth floor eliminates b_cross=0 divergence. gamma_conf_eff substitution eliminates gamma_conf bypass. Supercritical pitchfork reinterpretation resolves C9-5.
  • ·Hungary TISZA election resolved 2026-04-12: TISZA won in a landslide (Orbán conceded after 16 years). Our P=0.80, Brier=0.0400. Market P=0.73, Brier=0.0729. Formula beat market by 0.0329 Brier score — first win.
  • ·CROSS-013 downgrade REJECTED: Sub-Agent 5 claimed 'alpha>1 gives finite variance' — correct threshold is alpha>2. Sub-Agent 15 (methodology evaluator) failed to catch the error. Philosopher issued C11-4 mandatory QA audit for all future sessions.
Session 10April 11, 2026Partial approvalv0.5.4

Session 10: The Limit Cycle Discovery

Lead: Computational Sociologist

  • ·Hopf bifurcation insight: Turchin's 4 secular cycle phases (integration → stagnation → disintegration → depression) are temporal quadrants of a single limit cycle, not four separate attractors. Validated by Wittmann & Kuehn 2024 (PLOS ONE, 5/5) — strongest single source in the formula.
  • ·Period discrepancy resolved: three compounding nonlinear factors (Krylov-Bogoliubov amplitude dependence, elite coupling strength, institutional damping) explain why the 140yr linearized period appears as ~250yr empirical cycles.
  • ·Limit cycle phase assignment: 4/5 testable historical events correctly placed in the right quadrant, suggesting the dynamical structure is capturing something real.
Session 9April 11, 2026Rejected

Session 9: The Network Layer (Rejected)

Lead: Network Scientist

  • ·Social networks are NOT scale-free: Broido & Clauset 2019 — reclassified to truncated power-law (gamma_sf ~ 2.3). SIR/SIS epidemic models inappropriate for opinion spread (complex contagion requires ~25% threshold, not single exposure).
  • ·Misinformation spreads ~6x faster than truth: Vosoughi et al. 2018 (Science, 126K stories) — novelty drives differential, not bots.
  • ·G_t 5-tuple APPROVED with caveats: G_t = (P_k, C, L, Q, rho_e). A_8 connectivity dynamics APPROVED: A_8(kappa) = alpha_net * (kappa_eq - kappa).
Session 8April 10, 2026Approved with caveatsv0.5.3

Session 8: The Economic Power Laws

Lead: Econophysicist

  • ·Inverse cubic law alpha_tail ~ 3.0: 40M+ data points across multiple markets (Gopikrishnan 1999, Gabaix 2003, methodology 5/5). Confirms FP+jump split is correct: finite variance for continuous drift (alpha > 2), infinite variance for crisis jumps (alpha_war = 1.53 < 2).
  • ·lambda_labor ~ 0.57 from BLS/FRED data resolves OPEN blocking item: epsilon = (1-w)/(0.57*e) now computable, yielding ~44x income ratio for top 2% — consistent with empirical data.
  • ·CROSS-013 CRITICAL: alpha_pareto = 1.5 implies infinite variance for elite fraction e — the diffusion tensor D_ee cannot use the raw Pareto distribution. Requires tempered Pareto or alternative specification.
Session 7April 9, 2026Approved with caveatsv0.5.2

Session 7: Structural Repair

Lead: Statistical Physicist

  • ·CRIT-01 RESOLVED: Psi and phi removed as state dimensions — they are observables computed from S_t, not independent variables. State vector reduced from 10D to 8D. Markov property preserved. Prevents DAE numerical instability.
  • ·CRIT-02 PARTIALLY RESOLVED: Jump kernel K(S,S') = rho * omega specified. Magnitude: tempered truncated power law. Direction: Boltzmann-weighted basin bias toward lower-energy attractors. T_eff temperature parameter remains undefined.
  • ·CROSS-002 RESOLVED: gamma_conf_effective uses sigmoid saturation — prevents [0,1] overflow without clipping.
Session 6April 8, 2026Approved with caveatsv0.5.1

Session 6: The Genetic Floor

Lead: Evolutionary Psychologist

  • ·HYBRID model: Theta_total = Theta_fixed_floor + Theta_variable(culture,t). Most 'constants' have a genetic floor plus cultural overlay — resolves the micro/constants boundary dispute.
  • ·alpha_recip ~ Normal(0.45, 0.08): Spadaro et al. 2022 meta-analysis across 70 societies and 1,506 studies — 'very little cross-societal variation in impersonal cooperation.' Strongest cross-cultural validation in the formula.
  • ·theta_boundary fractal layers [5/15/50/150/500/1500]: scaling ratio ~3.0 validated across phone networks, online platforms, and face-to-face groups. The specific '150' is contested (Lindenfors 2021 CI [4-520]) but the ratio is robust.
Session 5April 7, 2026Approved with caveatsv0.5.0

Session 5: Institutions Enter the Formula

Lead: Political Scientist

  • ·I_t defined as 5D vector (R_t regime type, V_t veto players, B_t bureaucratic inertia, P_t propaganda/trust, X_t external constraints) with 8 per-equation drift modulations.
  • ·pi_path = 0.96 annual: institutions have extreme path dependence — half-life ~17 years. Colonial-era institutional quality explains R² > 0.25 of modern governance variation.
  • ·Philosopher quote: 'This session is a genuine scientific contribution. Jones & Olken's death-in-office instrument is one of the strongest natural experiments in political science.'
Session 4April 7, 2026Approved with caveatsv0.4.0

Session 4: The Confidence Architecture

Lead: Bayesian Statistician

  • ·Hard predictability ceiling: R² < 0.50 for aggregate social prediction (Martin et al. 2016, Science). Lyapunov time 5–20 years for macro-social dynamics — horizon beyond which predictions become noise.
  • ·Fat-tail constraint critical: alpha_war = 1.53 < 2 means the jump process has infinite variance. Standard confidence intervals DO NOT EXIST for discrete crisis events. Dual-likelihood architecture required: Gaussian for FP continuous component, truncated power-law for jump component.
  • ·Prior domination acknowledged: ~30 parameters vs ~8 independent historical observations. Posteriors are dominated by priors, not data — an honest but concerning disclosure.
Session 3April 6, 2026Approved with caveatsv0.3.0

Session 3: The Historical Ground Truth

Lead: Cliodynamicist

  • ·T_secular ~ 250yr validated: Turchin & Nefedov (2009) across 8 secular cycles in 4 countries. LogNormal distribution, CV ~0.25. This is the foundational timescale of the formula.
  • ·alpha_war = 1.53: Richardson's law for war frequency (Clauset 2018, 95 wars, KS p=0.78) — highest empirical confidence of any parameter. Implies infinite variance for crisis magnitudes (alpha < 2).
  • ·F_fiscal = sigmoid((d-0.8)/0.2): Reinhart-Rogoff 90% threshold debunked; continuous sigmoid used instead. Critical for PSI's SFD component.
Session 2April 6, 2026Approved with caveatsv0.2.0

Session 2: The Particle Parameters

Lead: Behavioral Neuroscientist

  • ·Loss aversion lambda ~ Normal(2.25, 0.65): Brown & Inman 2019 meta-analysis (84 samples). Replication status: MODERATE (several failures in financial domain).
  • ·Temporal discount beta_td ~ Beta(2, 5): Samuelson exponential as baseline; hyperbolic corrections documented. Large individual and cross-cultural variance.
  • ·Conformity gamma_conf ~ Beta(3, 4): Asch-type meta-analysis; varies systematically with group size and unanimity.
Session 1April 6, 2026Approved with caveatsv0.1.0

Session 1: The Mathematical Foundation

Lead: Statistical Physicist

  • ·Framework chosen: Fokker-Planck equation with jump process — proven in Weidlich (1971), Toscani (2006), Scheffer et al. (2009). Sociophysics is phenomenological modeling, not a category error.
  • ·State vector S_t defined as 10-dimensional: 5 directly measurable (n,w,e,d,U), 3 indirectly measurable (pi,T,kappa), 2 derived (Psi,phi).
  • ·Jump process tied to Turchin PSI composite: MMP × EMP × SFD. Base rate lambda_0 undefined; threshold Psi_threshold undefined — all retrodiction untestable.