Skip to content

Exploration 05 — RAT ↔ climate-lama convergence decision

Status: session 4c, drafted 2026-07-26 — RATIFIED 2026-07-26 with owner amendment (see Decided ledger): the products are simply distinct — no convergence question exists today. End-state vision: once climate-lama matures, RAT becomes one of many consumers of the climate-lama backbone/API, exactly like climate-lama-ui — they can even evolve in parallel. What crosses is gathered practice (especially Greek-market knowledge), nothing else. Why this memo exists: the reassessment's completeness critic found no recorded decision on the relationship between the two platforms, while the P1 validation channel, the Greek data assets (EFAS 93 m river flood, Storm Daniel EMSR697, ELSTAT boundaries), and the UI market scan all currently live in the other repository (risk-assessment-tool, "RAT"). Left undecided, the next phase either duplicates work blindly or entangles the two tracks.

The facts

RAT (risk-assessment-tool) climate-lama (this repo + engine + ui)
Track Corporate delivery/demo vehicle (insurer demo pending) Independent platform product (this exploration track)
Lineage Descendant of the riskwise line Clean-room, no shared git ancestry or code lineage
Shape Desktop-first application Web-first, API-first, multi-tenant platform
Greek data today EFAS 93 m RF, Storm Daniel, ELSTAT — ingested there Historical JRC RF (bbox demo) + LitPop only
Shared dependency Consumes climate-lama-engine (public PyPI, Apache-2.0) Owns it
  1. No code transfer in either direction, ever. Each codebase keeps clean, independent provenance. The only shared artifact is climate-lama-engine, which is already a public, versioned PyPI dependency both may consume — that is the seam, and the only seam, at the code level.
  2. Data: same public sources, separate pipelines. climate-lama ingests Greek open datasets (EFAS/Copernicus, Storm Daniel EMSR697, ELSTAT, GADM, ESRM20) independently from their original public sources, through its own ingest pipeline with its own attribution records — never by copying artifacts, configs, or processed outputs from RAT. Slower, but every dataset in climate-lama carries self-contained provenance.
  3. Knowledge flows freely; deliverables don't. Market-scan facts, UI patterns (the ui-revamp study), methodology reading, and interview learnings are facts and skills, not work products — they inform both tracks. Demo assets, reports, client-specific configurations do not cross.
  4. Validation signals cross as signals. The corporate track's insurer feedback feeds climate-lama's checkpoint gates (P1 conditionality, ORSA interview checkpoint) as evidence about the market, nothing more. climate-lama never depends on RAT deliverables to build or demo — its own demo path stays self-contained (compose fullstack + seeded data).
  5. No roadmap coupling. Neither track blocks the other; this track's phase definition includes its own Greek-data ingest items (the scenario data pack, session 5) rather than assuming RAT's data work transfers.

Why this split (and not convergence)

  • Provenance is the asset. climate-lama's value as an independent work rests on its clean lineage; the cheapest way to destroy that is casual cross-repo reuse.
  • The tracks have different owners of record and different fates — decided in rooms this repo has no visibility into. Defined seams mean no outcome of either track contaminates the other.
  • Duplication cost is bounded and mostly one-time (re-ingesting open datasets through an existing pipeline), while entanglement cost is unbounded.

Decided / Parked / Killed (session 4c)

Decided (ratified 2026-07-26) - The five-rule seam model above, with the owner's framing amendment: these are two distinct products — RAT today is close kin to its riskwise-v2 predecessor; the mature climate-lama end-state has RAT as one of many API consumers (peer to climate-lama-ui), possibly developed in parallel. Gathered practices (Greek-market knowledge above all) transfer; nothing else does.

Parked - The timing/mechanics of RAT consuming the climate-lama API — a future decision for when the platform is mature; nothing in this track blocks or depends on it.

Killed - Implicit convergence ("the Greek data is basically ours already") — every dataset enters through this repo's own pipeline or not at all.