Mind · Inference
TemporalEnsembleForecasting
Forecast by ensemble of model-checkpoints from different history horizons, each run through five non-substitutable causal lenses — convergence across time is the signal
Definition
Mechanism
A protocol for forecasting via ensembles of models frozen at different points in history (Temporal Checkpoint Ensemble). Decomposes prediction via ConceptualDecomposition along two axes simultaneously: temporal diversity (each checkpoint reasons from a different knowledge horizon — the 2020 checkpoint knows nothing of COVID's economic aftermath; the 2022 checkpoint has internalized it) and causal-lens diversity (each forecast is run through the Forecasting Pentagon: structural, economic, political, base-rate, temporal — five non-substitutable angles that suppress each other if entangled in a single pass). A ReduceSolver composes the matrix of forecasts, not by averaging but by preserving causal reasoning from each horizon and identifying where independent knowledge states converge on the same structural driver. When three checkpoints spanning different knowledge states independently identify the same causal mechanism, that mechanism has survived a temporal adversarial test no single-model forecast can replicate.
Invariants
- Temporal diversity source: the checkpoints differ by training-data cutoff, not by prompting — history itself is the diversity generator.
- Five-angle Forecasting Pentagon: each checkpoint's forecast is decomposed across structural, economic, political, base-rate, and temporal dimensions (the pentagon passes the four-test decomposition).
- Convergence-not-averaging: the reduction gate accepts convergence of structural drivers across horizons, not arithmetic average of point predictions.
Risks
Failure modes
None stated.
Relationships
Builds on
Composes with
Finding what a concept is made of: locate it, carve it into the dimensions every case it claims needs, test each, search for what is missing, route real cases through it, say how the parts make the whole, and carve again below
Nothing in this library builds on this pattern yet.