Probability components
The live model combines a structural-risk score (v4 = isotonic(kio_hist_5y + 2 × kio_recent_7d)) with the OONI-trajectory forecast (honest_forecast_v1); the served version is shown above. The logit-space blend (sigmoid(0.6·logit(v4)+0.4·logit(hf))) is auto-gated — it serves only on days it durably beats the live model, leave-one-country-out validated.
This score ranks which countries are heading into danger (forward country-ranking AUC 0.89 against Access Now KeepItOn labels, measured in the committed model-stats snapshot). It is not a within-country timing signal: in a strict past→future holdout, within-country timing AUC is 0.45, below chance. Live numbers: /v1/shutdown-risk/info.
- ensemble_raw
- 0.001
- hf_prob
- 1.0%
- sr_prob
- 0.0%
Honest validation
Same validation block this endpoint returns inline — never computed at request time, never adjusted per-country.
Model honesty caveats (7)
- It mostly ranks countries by how often they have had shutdowns. It has not been shown to predict when a new shutdown will start; within-country timing is unvalidated (B1600).
- The served version can change daily: an automatic gate chooses it. Read model_version before citing anything.
- In the strict past->future holdout (isotonic fitted on the first 75% of dates, scored on the rest): cross-country AUC 0.909, country-ranking AUC 0.943, PR lift 8.53x over the base rate; within-country median AUC 0.382 over 13 countries, below chance (0.5). The cross-country figures compare countries; they do not time a shutdown within a country. The in-sample within-country median is cross-sectional (full panel, in-sample isotonic), not forward skill.
- v9 replaces v5's multiplicative product with a logit-additive blend (alpha=0.6 on structural risk). The product collapsed to ~0 when honest_forecast was near-zero for a country, annihilating a good structural signal; the logit blend cannot zero out either factor.
- In-sample within-country median AUC 0.593 (full panel, in-sample isotonic; not forward skill), lower than the 0.729 cited for v5 before: a drop, not an improvement. In the strict past->future holdout the v9 within-country median AUC is 0.382 over 13 countries, below chance (0.5); the forward cross-country AUC is 0.909. Leave-one-country-out selection only chose the blend weight (alpha=0.60); it does not show within-country timing skill.
- Still country-level; sub-national shutdowns (~half of KeepItOn events) remain harder. Low-N countries are sample-size limited, not model-limited.
- Same two honestly-validated base models as v5 (v4 structural + honest_forecast_v1). No new external features (unlike the gated v7 CF-Radar and v8 GDELT attempts) — purely a better combiner.
About this view
Experimental 7-day internet-shutdown risk score from shutdown_risk_v9; onset timing is unproven. Trained on and compared against Access Now KeepItOn STOP shutdown records (journalist-verified). Live JSON: /v1/shutdown-risk/DE.