AI & Architecture #

Solution Paradigm Decision Tree #

Purpose: route a given decision step to the cheapest paradigm that actually solves it. Bias is deliberately toward the lower-cost, lower-autonomy answer; escalation must be argued for, not assumed.

Rules #

Rule 1 — The unit of analysis is a decision step, not a system. #

“Claims triage” is not a tree input. “Determine whether the submitted document is an FNOL, a repair estimate, or correspondence” is. A decision step has a nameable input, a nameable output, and a stateable criterion for correct. If you cannot state all three, you are not ready to enter the tree.

Rule 2 — The output is a decomposition, not a label. #

A realistic system exits the tree five or six times with five or six different answers. If your system came out with one label, you didn’t decompose it.

Rule 3 — Run gates in order. Do not skip ahead. #

The gates are ordered by cost. Skipping to Gate 4 because the request said “AI” is the single most common failure of this tree.

Rule 4 — Composition is a first-class output. #

When you’re done, you have a map of steps → paradigms, plus the glue. The glue is almost always deterministic code, and it is where most of the actual engineering lives.

The Tree #

DECISION STEP
│
├─ GATE 0 ── Should this step exist at all?
│            ├─ no  → DELETE / REDESIGN THE PROCESS
│            ├─ no  → HUMAN DOES IT
│            └─ yes ↓
│
├─ GATE 1 ── Can a competent human write down the rule?
│            └─ yes → [A] DETERMINISTIC LOGIC
│                     (code · decision tables · rules engine · DMN)
│            └─ no  ↓
│
├─ GATE 2 ── Is this &#34;choose the best combination under constraints&#34;?
│            └─ yes → [B] OPTIMISATION / SEARCH
│                     (MIP · CP-SAT · heuristics · simulation)
│            └─ no  ↓
│
├─ GATE 3 ── Structured/historical inputs + labelled outcomes?
│            └─ yes → [C] LEARNED MODEL
│                     ├─ C1 statistical / actuarial (GLM, GAM, credibility)
│                     ├─ C2 classic ML (GBM, RF, ranking, clustering)
│                     └─ C3 fine-tuned / small specialised model
│            └─ no  ↓
│
└─ GATE 4 ── FOUNDATION-MODEL LADDER — start at rung 0, justify every step up
             ├─ R0 retrieval only          — find, rank, surface. No generation.
             ├─ R1 single call             — one bounded call inside your code.
             ├─ R2 bounded workflow        — many calls, YOU wrote the sequence.
             ├─ R3 constrained agent       — model picks order, fixed toolset, capped.
             └─ R4 autonomous agent        — open objective, open toolset, long horizon.