Limits & Failure Modes

This page documents known limitations and failure modes of the Global Curiosity Engine (GCE) protocol.
These are limitations of the method, not of the domains to which it is applied. Listing them
explicitly is a core requirement of the protocol itself.

Scope limitations

• Trend-only sensitivity (v1.0):
  GCE v1.0 is limited to detecting instability in directional trends. It does not assess magnitude,
  volatility, causal structure, or distributional shape beyond what is required to estimate direction.
• Long-horizon focus:
  The protocol is designed for multi-decade or long-run baselines. It is not appropriate for
  short-term forecasting, event detection, or real-time monitoring.

Data-related failure modes

• Aggregation masking:
  Spatial or categorical aggregation (e.g., national averages) can smooth or hide sub-regional
  instability. GCE does not resolve within-unit heterogeneity unless explicitly applied at finer resolution.
• Temporal smoothing:
  Use of multi-year or averaged series may suppress short-lived shocks or regime transitions.
  Apparent stability may reflect smoothing rather than genuine continuity.
• Revision artefacts:
  Dataset revisions may introduce apparent instability unrelated to the underlying phenomenon.
  Cross-revision checks mitigate but do not eliminate this risk.

Parameter sensitivity

• Threshold dependence:
  Classification of “increase,” “decrease,” or “stall” depends on explicit thresholds. Different
  thresholds can introduce or remove marginal reversals. Sensitivity must be reported, not optimized away.
• Window choice:
  Rolling window length affects detection. Windows that straddle structural transitions may amplify
  or dampen apparent reversals. Persistence across window lengths is required for PSZ confirmation.

Interpretive risks

• False epistemic certainty:
  A PSZ indicates predictive fragility, not factual error. Treating PSZ as proof of model failure
  or domain collapse is a misuse of the protocol.
• Narrative substitution:
  There is a risk that users replace discarded certainty with speculative explanation. GCE explicitly
  forbids this; interpretation must remain external to the protocol.

What GCE cannot tell you

• Why a trend changed.
• What will happen next.
• What actions should be taken.
• Which interpretation is “correct.”

Design intent

These limitations are not flaws to be patched. They are boundary conditions that preserve the
epistemic integrity of the protocol. GCE is designed to fail early, visibly, and conservatively—
preferring false negatives over false confidence.

Page last updated: February 2026