Engine state: Moderate risk / Evaluate (Yellow) — saturated ground, mean forecast still modest
Nairobi admin-1, early March long-rains onset. Dense informal settlements along the Nairobi/Mathare/Ngong river corridors are exposed to flash flooding. Ground is already wet from late-February rain. You are the DOC duty analyst on Mar 1.
Act I quiz — build the evidence inventory Every answer becomes an input to the risk model: evidence → classification → confidence → belief update → decision. Not scored.
Q1. The rainfall outlook for Nairobi comes from the ECMWF ensemble (many parallel model runs, updated daily). First, the basics: what is a DETERMINISTIC forecast? A single model run giving one outcome — no range of possibilities Many model runs combined into probabilities A forecast that is always correct An observation of what has already happened BN purpose: Forecast literacy — why a single number is not enough for risk decisions.
Q2. And what is an ENSEMBLE prediction system (EPS)? The same model run many times from slightly different starting conditions, giving a spread of outcomes Several forecasters voting on the most likely outcome One model run at a higher resolution An average of past observed seasons BN purpose: Why forecasts are SOFT evidence in the BN — the spread is the confidence.
Q3. What is the strongest evidence currently available? On your desk this round: 7-day antecedent rainfall (IMERG) · Ensemble-mean exceedance prob (ECMWF P_heavy)
Forecast Observation Community report Historical analogue BN purpose: Determines which BN node receives the highest weight.
Q4. Classify that evidence. Check the tags on the evidence cards: 7-day antecedent rainfall (IMERG) · Ensemble-mean exceedance prob (ECMWF P_heavy)
Hard Soft Virtual BN purpose: Determines confidence weighting — hard = observation (one-hot), soft = probability vector, virtual = likelihood message.
Q5. How reliable is this evidence? Nairobi is a TAIL-risk case — the ensemble mean can look benign while a few members carry the danger; weigh the tail, not the average. Reliability sets how hard the BN leans on it.
Very Low Low Moderate High Very High BN purpose: Evidence likelihood weighting.
Q6. What hazard condition does this evidence support in Nairobi, Kenya? Which mechanism fits Nairobi — slow river rise, sudden flash runoff, or urban-drainage overload? The mechanism decides which impacts follow.
Heavy rainfall River flooding Flash flooding No hazard BN purpose: Updates the hazard node.
Q7. Which impact pathway is most likely here? Nairobi admin-1, early March long-rains onset. Dense informal settlements along the Nairobi/Mathare/Ngong river corridors are exposed to flash flooding. Ground is already wet from late-February rain. You are the DOC duty analyst on Mar 1.
Population displacement Road disruption Crop loss Water contamination BN purpose: Updates the impact node.
Q8. What evidence would you request next? Still off the table this round: Ensemble-max / 2-yr RP, pixel p95 · Spatial coverage of heavy-rain pixels · Yesterday's risk posterior carried forward (DBN, alpha=0.6) · Observed rainfall / river response. Which would most reduce your uncertainty before you decide?
Satellite observations Gauge observations Hydrological model Field assessment Exposure information BN purpose: Real DOC operations seek more evidence — watch which cards the Act II rounds reveal.
Q9. On 1 March 2026 the 7-day antecedent rainfall over Nairobi is about 135 mm and the ground is saturated, while the ensemble-mean forecast still looks modest. Why does the antecedent evidence alone justify stepping up vigilance? The brief notes the ground is already wet from late-February rain.
It does not; a modest mean forecast means low risk regardless of soil state Saturated ground converts almost any further intense storm directly into runoff, so the same forecast rain is far more dangerous than it would be on dry ground Wet soil mainly increases drought risk later in the season Antecedent rainfall only matters for large rural river basins, not cities BN purpose: Why the antecedent_rainfall node shifts the flood prior before any forecast node fires.
Q10. At T-2 (4 March 2026) the ensemble mean is only about 18 mm, but the worst ensemble member gives 131 mm and the p95 ensemble-max ratio crosses the 2-year return period (1.11). What is the correct reading of this split? Mean and tail are answering different questions.
The mean is the best estimate, so the heavy members can be discarded as noise A minority of members produce an extreme convective storm: the mean hides a real tail risk that a mean-threshold system would miss, so precautionary action at 2-day lead is justified The disagreement means the forecast is unusable and no action should be taken The tail member guarantees a flood will occur on 6 March BN purpose: Why the tail_risk node, not exceedance_prob alone, is the decisive evidence for tail-forecastability events.
Q11. Your current risk estimate — before the model runs? Nairobi is a TAIL-risk case — the ensemble mean can look benign while a few members carry the danger; weigh the tail, not the average. No model has spoken yet — commit your own read so Act III can compare it to the engine.
Low Moderate High Very High BN purpose: Your pre-BN estimate. Act III compares it against the engine risk indication.
Q12. Recommended DOC status — commit now, before seeing the BN output. At T-2 (Mar 4) the ensemble MEAN is benign but the tail crosses the 2-yr RP. What DOC level do you set, and why? A 'Watch + 24h re-check + note the tail uncertainty' is a strong answer.
Monitor Watch Warning Emergency Coordination BN purpose: Your pre-BN decision. Act III compares it against the CRMA state.
Q13. The risk model you are about to work with was built from expert rules, not calibrated against decades of events. How much should you trust it? No right answer — your stance is revisited in Act III against what the model actually did, and what history recorded.
Fully — it computes probabilities Not at all — rules are subjective As a transparent colleague — consistent and inspectable, but encoded judgment BN purpose: Hold this question. Act III returns to it.