Limitations of this system

Early-warning tools earn trust by being clear about their edges. Every honesty flag elsewhere in this app — the Live / Simulated label on each data card, the “rule-based, not AI” framing of the score, the illustrative labelling on the backtest — points back to this page.

1. 1. Rainfall and dam status are not the only drivers of flooding

   The score is built from measured rainfall against basin thresholds, compounded by reported dam capacity. Real flooding also depends on soil saturation and antecedent moisture, upstream tributary inflow from catchments outside the modelled basin, river channel condition and siltation, drainage and urban surface sealing, and infrastructure beyond the dams we track — culverts, embankments, small private dams. None of those are captured here. A low score does not mean a basin is safe.

2. 2. It depends on live API uptime, and degrades honestly rather than silently

   Live rainfall comes from the Open-Meteo public API on an hourly schedule. If that feed is unavailable, scores go stale rather than being quietly interpolated: the last reading keeps its own timestamp, and every card states whether the data behind it is Live or Simulated. Offline citizen reports queue locally and sync when connectivity returns, with a visible banner while queued. We would rather show you a gap than fill it with a guess.

3. 3. Thresholds are seeded estimates, not peer-reviewed hydrology

   Each basin's 24-hour and 7-day rainfall thresholds, the 50/50 point split, and the dam compounding weights (+10 near capacity, +25 overflowing) were chosen to produce sensible, explainable behaviour for this system. They have not been calibrated against gauge archives or reviewed by hydrologists, and they are not derived from a published rainfall-runoff model. Before any operational use they would need validation against official KMD and WRA records, per basin.

4. 4. The backtest is illustrative analysis, not a certified validation study

   The multi-year backtest reconstructs the reported shape of the March–May 2024 and March–May 2018 long-rains seasons — a steady build, a peak, a recession — rather than replaying gauge records. Response timings are approximate readings of published reporting. The resulting lead-time figure is an illustration of how the rule behaves on that shape, not evidence of forecast skill.

5. 5. Human judgment from trained responders remains essential

   This is decision support, not a decision maker. It is rule-based, not AI, and it does not know about road closures, local evacuation capacity, ongoing operations, security constraints, or the hundred contextual factors a county disaster officer holds in their head. Suggested actions are prompts to consider, not instructions to follow. Nothing here should override the judgment of trained emergency responders, and no one should wait for this system to turn red before acting on what they can see.