Pick the right method,
then meet the human who can supervise you.
A guided walk-through. The researcher describes their question and their data. M12 navigates a structured decision tree — question type, data structure, spatial/temporal/network dependence — backed by Claude, and proposes a methods family with starter-code references. Two outcomes every session: a one-page methods brief and a routing recommendation to a named adviser in the working-modeller network.
Question typology
Descriptive, predictive, causal, optimisation, mechanism-informed inference. Each branches to a different methods family. No guessing.
Data-structure check
Point-level / area-level, time-series / panel, network. The data structure constrains the methods. The Coach forces you to look at it before recommending anything.
Methods family + starter code
Bayesian hierarchical, GP-based spatial, time-series state-space, mixed-integer optimisation. Each comes with a link to the AMB starter-pack template repo.
Trap-flagging
Three Bayesian-modelling traps every African modelling project hits — NegBin replacing the GP, case-only likelihood claiming identifiability, 3D space-time product kernels. M12 flags them up-front.
Adviser routing
Every Coach session ends with a recommendation to a named working-modeller adviser (Nick / IDEM-LAB / MAP / Vector Atlas / AIMS / ACEFA) — chosen on methods match, not on availability.
One-page brief
The session output is a downloadable methods brief you can hand to your supervisor, the funder, or the assigned working-modeller adviser in the next dialogue session.
M12 is intentionally conservative. The decision tree is hand-curated from the working-modeller advisers' guidance and only the explanatory copy is Claude-generated. The model never recommends a method outside the curated set. The final outcome is a *human* — not a model — to ask. M12 always recommends the named adviser as the next step, not itself.
The live module in v0.0 is M10 · Document Triage.
A working demo of the same back-end architecture. Try it.