# Taiwan Election Observatory — Validation Report v0.1

## Current status

**Predictive validation: not evaluated.** The number of independent election cycles evaluated is zero in this first public release. Synthetic tests establish mechanical correctness only; they do not establish forecasting accuracy.

## Mechanical checks implemented

- Candidate shares sum to 100 percent in every draw.
- Winner probabilities sum to one and use nationwide plurality.
- Fixed seeds reproduce identical summaries.
- Research mode returns `insufficient-evidence` instead of synthetic predictions.
- Polls after the cutoff and incompatible candidate configurations are excluded.
- Baseline and scenario outputs are separate objects.
- X uses complete-distribution mixture sampling with normalised nonnegative weights.
- Import validation rejects incompatible configurations and missing availability timestamps.
- Saved snapshots record cutoff, configuration, version, seed, weights, checksums, assumptions, results and evidence status.

## Historical evaluation pathway

When archived inputs are available with reliable public-availability timestamps, replay approximately 180, 90, 30 and 7 days before each election. Use chronological election-cycle holdouts rather than random weekly splits. Compare historical reference, A, B, C with and without events, and X using vote-share MAE/RMSE, interval coverage and width, multiclass Brier score, log score and sensitivity analysis.

## Remaining research gaps

- Compatible named-candidate poll archive with wording and disclosure metadata.
- Economic data vintages known as of each replay cutoff.
- Sufficient independent presidential election cycles for restrained fundamentals fitting.
- Consistent event coding with revision and cluster history.
- Prospective tracking under a confirmed candidate configuration.