Colombia 2026 presidential runoff forecast — v2.0.1 implementation

**Implementation version v2.0.1** of the Colombia 2026 forecasting platform. A narrow R1 reconciliation of the v2.0.0 deposit ([10.5281/zenodo.20618031](https://doi.org/10.5281/zenodo.20618031)). The R2 runoff posterior carries forward bit-identically; only the R1 first-round posterior and its calibration audit changed. Pre-registered methodology: OSF [osf.io/3ug2m](https://osf.io/3ug2m) (immutable); v1.0.0 Zenodo deposit [10.5281/zenodo.20578542](https://doi.org/10.5281/zenodo.20578542). **R2 forecast (carried forward bit-identically from v2.0.0):** - Cepeda voter share: 47.57 % [46.75, 48.41]- Espriella voter share: 52.43 % [51.59, 53.25]- Margin (Cep − Esp): −4.85 pp [−6.50, −3.18]- P(Espriella wins): 1.000- Abstention (eligible): 43.68 %- `runoff_tier2.nc` SHA-256: `f41674b722b734afe4bb78fdabf80e43150c24eb7a1736868c2fc4b95ce9c949` (byte-identical to the v2.0.0 deposit) **v2.0.1 changes relative to v2.0.0:** 1. **R1 posterior reconstruction.** Refit with a historically-calibrated populist-underestimation prior: `mu_pop = -0.25` for right-populist candidates (de-la-Espriella, Valencia), calibrated from 2018 R1 Duque (actual 39.1 % vs. polls ~36 %, +3 pp miss) and 2022 R1 Hernández (actual 28.2 % vs. polls ~20 %, +8 pp miss). The negative sign reflects the documented Colombian polling-miss direction for right-populist candidates; `populist_under` is added to the predicted poll logit, so negative values shift inferred truth upward relative to the polled signal. Late-shock prior reverted to Normal (v2.0.0 used Student-t(ν=3) per the Item G amendment). R1 fit frame reverted to `mrp_frame_v0.5.0.parquet` to align with the v1.0.0 baseline.2. **R1 fit profile.** Research-heavy NUTS (4 chains × 2 000 warmup × 4 000 sampling = 16 000 draws). Layer-1 PASS: R-hat 1.0027, ESS bulk 9 363.88, ESS tail 5 782.86, divergences 0, E-BFMI 0.95.3. **R1 calibration audit refresh.** `r1_model_vs_actual_v1.0.0.json` regenerated against the refit posterior. Overall MAE 11.06 pp; 80 % credible-interval coverage 35.8 %; 165 (dept × candidate) pairs. National rollup (eligible-weighted, share among 5 modeled candidates): Cepeda 32.12 % (actual 43.31 %), Espriella 33.82 % (actual 44.74 %), Fajardo 12.06 %, López 9.20 %, Valencia 12.80 %. Predicted R1 finalist margin Esp +1.70 pp, matching the actual R1 margin (Esp +1.43 pp) within 0.27 pp.4. **Receipt UTF-8 re-encoding.** Receipt JSON files re-encoded as UTF-8 (the v2.0.0 deposit's receipts used Windows cp1252 for the § byte). **Binding commits:** - R2 runoff posterior: git tag `v2.0.0` → commit `20f2831a355e0fa3d70aa5eff21db83e0a78a11b`- R1 first-round posterior: git tag `v2.0.1` → commit `e707689586f22730af9c58bd5f08428db50ea3a9` **Validation outputs (R2 carried forward from v2.0.0):** - Layer 1 (NUTS convergence) — PASS: R-hat 1.0021, ESS bulk 3 349.87, ESS tail 2 814.08, divergences 0, E-BFMI 1.95.- Layer 2c (runoff Brier on 2018 R2 + 2022 R2) — Tier 1 Brier 0.137 / 0.159; both exceed 0.07 threshold; Tier 2 used for production.- Layer 3 (Simulation-Based Calibration) — KS p-value 0.84 (uniformity holds).- Layer 4 (single-ablation LOO) — Item C is most load-bearing (ΔMAE 0.029).- R1 calibration audit (model vs. actual, eligible-weighted) — MAE 11.06 pp; 80 % CI coverage 35.8 % (v2.0.1 refresh). **Independent peer review (carried forward from v2.0.0):** Gemini 2.5 Pro, 2026-06-09. Verdict: APPROVED-WITH-FIXES. All 5 findings closed (1 CRITICAL TLS bundling, 2 IMPORTANT [PDF fixture tests + CNE walker max-pages], 2 MINOR [Datexco citation + LAPOP imputation transparency]). **Read first (top-level files in the tarball):** `README.md`, `IMPLEMENTATION_VERSIONS.md` (parameter catalog for v1.0.0 / v2.0.0 / v2.0.1), `BINDING_INTEGRITY.md` (reproduction instructions), `methodology_paper_v2.0.1.md`, `FINAL_FORECAST_2026-06-21.md`, `OSF_IMPLEMENTATION_NOTE_v2.0.1_draft.md`. **Tests:** 1172 / 1172 unit tests pass. **Citation:** Santiago, S. (2026). *Colombia 2026 presidential runoff forecast — v2.0.1 implementation*. Zenodo. doi: this record. Filed as a new version of the v2.0.0 deposit (10.5281/zenodo.20618031); same DOI lineage as v1.0.0 (10.5281/zenodo.20578542). Pre-registered methodology: OSF osf.io/3ug2m.