Validation

This page documents applications of the Global Curiosity Engine (GCE) protocol to public datasets. The purpose is
to validate the instrument’s behavior: whether candidate Predictive Stress Zones (PSZ) persist after falsification
(window perturbation, threshold sensitivity, and cross-revision or cross-dataset comparison). These runs do not
interpret causes, produce forecasts, or recommend actions.

Validation scope

  • Protocol version: GCE v1.0 (primary stress signal: trend instability)
  • Stress signal: directional trend reversal across rolling baselines
  • Falsification gates: window perturbation, threshold sensitivity, and cross-revision / cross-dataset checks
  • Outputs: PSZ listings, persistence summaries, method limitations, and question-only curiosity prompts

Run A — NOAA climate baselines (U.S.)

The protocol was applied to NOAA historical climate data using rolling 30-year baselines and simple, explainable
metrics. Candidate zones were flagged where directional trends reversed between successive baselines. Falsification
checks were performed to test whether flagged reversals persisted under perturbations.

Observed instrument behavior

  • Primary signal persistence: Trend-instability candidates persisted across multiple state–time windows.
  • Discard discipline: Weaker or brittle signals were not promoted without persistence under falsification.
  • Scope control: Outputs remained procedural (no climate causation, policy, or forecasting).

Recorded limitations (method, not domain)

  • State-level aggregation can smooth local extremes and mask sub-regional dynamics.
  • Heuristic thresholds are used for reversals; sensitivity is documented rather than hidden.
  • Confidence and uncertainty measures depend on available metadata and may be approximated when not provided.

Run B — UN World Population Prospects (WPP): Total Fertility Rate (TFR)

The protocol was applied to UN WPP total fertility rate (TFR) series using rolling 20- and 30-year baselines.
Directional trends were estimated for each baseline, and locations where the direction reversed were flagged.
Falsification was performed via window perturbation, threshold sensitivity, and cross-revision comparison.

Cross-revision check

  • WPP 2024 vs WPP 2022: Directional reversals were compared across revisions.
  • Signals that appeared only in one revision were downgraded or discarded.

Representative PSZ examples (illustrative)

  • United States: decline → modest rise → renewed decline (multiple reversals; persistent under perturbation).
  • Sweden: alternating declines and partial recoveries (high instability across decades).
  • France: long decline followed by modest upturn (single persistent reversal).
  • Kenya / Bangladesh: early increase followed by sustained decline (single persistent reversal).

Run C — Ultra-low fertility pass (v2.1): rebound assumption stress-test

A focused pass tested the stability of the “fertility rebound” assumption in selected ultra-low fertility contexts
using UN WPP data. The protocol remained unchanged: it flagged directional instability and then attempted to break
those signals using falsification gates. Persistent monotonic decline was treated as a high-salience form of
predictive fragility when rebound assumptions failed to hold under perturbation.

Observed instrument behavior

  • High persistence: In several ultra-low contexts, monotonic declines persisted under all perturbations.
  • Nuanced downgrades: Mixed or revision-sensitive candidates were recorded as weaker rather than forced into a binary.
  • Good collapse behavior: Some candidate signals weakened or collapsed under stricter thresholds/windows and were demoted accordingly.

Recorded limitations (method, not domain)

  • Temporal resolution (annual vs 5-year series) can change apparent instability; smoothing may mask shocks.
  • Stall vs decrease classification depends on explicit thresholds; sensitivity must be reported.
  • Successive WPP revisions can include methodological shifts; cross-revision checks are necessary but not sufficient.

What was learned (instrument-level only)

  • GCE reliably detects trend-instability candidates and distinguishes persistent signals from brittle artifacts.
  • Falsification gates function as intended: surviving PSZ are method-invariant under perturbation.
  • The protocol remains domain-agnostic when prohibitions are enforced (no causation, no forecasting, no prescriptions).

Curiosity prompts (question-only)

For each surviving PSZ, the protocol generates concise questions that target predictive assumptions. These prompts
are designed to guide human investigation without smuggling in explanations. (See domain runs for full prompt lists.)

Page last updated: February 2026

Scroll to Top