Climawise

How the Tool Generated This Brief

RecommendationsCases

The brief is produced by an evidence-retrieval and synthesis pipeline. Every step below is auditable. For a visual overview, view the flowchart.

  1. 1
    Build the risk-and-context profiles. Before any query, the tool turns raw data into a comparable profile for the target province and for every case-study location: climate (rainfall, temperature), terrain (slope, elevation), land use (cropland, built-up area, tree cover), hazard signals (flood and drought exposure, return periods), and population. Each indicator is restated on a shared z-score scale so different units can be compared fairly.
  2. 2
    Read your selection. The tool loads the selected province’s profile as the comparison baseline.
  3. 3
    Compose the province’s adaptation-need text. From the selected province’s profile and hazard, the tool writes a short natural-language statement of what adaptation the place needs: its dominant climate risks, its terrain and land-use context, and the kinds of measures that fit. This turns the structured profile into a query that describes the problem in words, so it can be matched against how case studies describe their own work.
  4. 4
    Embed the descriptive text (province and case study). A text-embedding model converts that adaptation-need text, and each case study’s described actions and outcomes, into numeric vectors. Comparing the vectors captures how relevant a case’s approach is to the province in meaning, beyond the structured indicators above.
  5. 5
    Find comparable adaptation cases. The case-study database is searched for projects that addressed the same hazard. Flood is split into two distinct hazards so the evidence stays on-topic: Riverine Flood draws on river and rainfall flood measures (drainage, storage, river training), while Coastal Flooding & SLR draws on coastal-water measures (seawalls, mangrove and dune defences, surge early warning) covering coastal inundation, storm surge, and chronic sea-level rise. Cases that cannot physically apply (for example, coastal flooding in an inland province) are filtered out before scoring.
  6. 6
    Score similarity on three tracks. Each case is scored on three kinds of similarity: how closely its measured risk and context match this province (from the profiles), how relevant its described actions are to this hazard (from the text embeddings), and how alike its geographic and implementation setting is. The three tracks combine into one composite score, and cases too different to be useful (weak or anti-peers) are set aside before selection.
  7. 7
    Group cases into archetypes (clustering). Candidate cases are assigned to archetype clusters from their feature profiles, so the engine can recognise when several candidates are variants of the same approach (for example, many districts of one national programme) rather than independent pieces of evidence.
  8. 8
    Select a diverse shortlist (MMR). The shortlist is built with Maximal Marginal Relevance: the engine adds cases one at a time, each time choosing the candidate that best balances high similarity to the province against being different from those already picked (weighted roughly 0.7 relevance to 0.3 diversity). A cluster-aware penalty can further discourage adding another case from an archetype already represented, so the final set spans distinct approaches instead of near-duplicates. The selected cases are then split into close peers and transferable examples.
  9. 9
    Synthesize recurring actions. The selected cases are passed to a language-model synthesis step (gpt-5.5, via the OpenAI Responses API) that aggregates adaptation actions across the set, grouped by IPCC AR6 lever: reduce hazards, reduce exposure, reduce vulnerability. Each drafted action must cite at least two supporting cases from the set.
  10. 10
    Check feasibility and confidence. Each drafted action is validated against the province profile (coastal vs inland, terrain, urban share, the queried hazard’s local signal) and against the feasibility rule set. A confidence label is assigned from the evidence count, the supporting cases’ quality, the regional reach of those cases, and the rule outcomes. Confidence is capped when the queried hazard’s local signal is below the national lower quartile, and recommendations the tool cannot defend are dropped.
  11. 11
    Multi-hazard handling. Two modes are available. Single hazard ranks against one hazard. Joint scoring ranks against a validated hazard combination (for example Riverine Flood + Drought, or Coastal Flooding & SLR + Tropical Cyclone): cases are scored over the unified feature subspace of the member hazards, and only cases that address the whole combination are shown. Each joint case carries a per-hazard contribution breakdown so the combined score stays auditable. Combinations outside the validated set are not offered.