CSE Roadmap

Common tier · about 5 hours · 13 parts

Part 2 of 13 · 2 min read

Filter by hard requirements and drop dominated options before any weighting

A trade-off only exists among options that already clear the bar; weighing a disqualified option is a waste of an argument.

Think of apartment hunting. Before comparing rent against commute time, you first drop every listing outside your budget. You also drop every listing worse than another one on every single feature. Neither kind of listing belongs in the comparison at all.

In plain words: a hard requirement (constraint) is a limit an option must meet to be considered at all, and a dominated option is one that another option matches or beats on every criterion and beats outright on at least one -- removing both before any weighting means the eventual weighted comparison is only ever deciding between options that were already worth comparing.

A worked example. For five candidate architectures of a support-agent design, applying two hard requirements -- p95 latency at most 8 seconds, quality at least 0.72 -- removes 3 of 5 options immediately, before a single weight is assigned. Of the two survivors, neither dominates the other: the small-planner variant is faster and cheaper, the small-planner-plus-cap variant has slightly better quality, so the comparison that follows is a genuine trade-off between two options that are each better on a different axis.

The key lines. From code/mini-tradeoff/matrix.py:

# from code/mini-tradeoff/matrix.py
def apply_constraints(options, constraints):
    """constraints: {criterion: (op, limit)} with op in '<=', '>='. Options that break a hard requirement are removed
    before any weighting: a trade-off only exists among options that are acceptable."""
    keep, dropped = {}, {}
    for n, v in options.items():
        bad = [c for c, (op, lim) in constraints.items() if (v[c] > lim if op == "<=" else v[c] < lim)]
        (dropped if bad else keep)[n] = bad
    return {n: options[n] for n in keep}, dropped

Read it as: the docstring states the reason for the function's existence directly. "A trade-off only exists among options that are acceptable" is the whole argument for doing this step first. Weighting criteria is a way to choose among acceptable options. It is not a way to decide whether an option is acceptable in the first place.

Hard requirement, Pareto-dominated

A hard requirement (constraint) is a limit an option must meet to be considered at all, such as a latency ceiling or a privacy rule. A dominated option is one where another option is at least as good on every criterion and better on one; a dominated option should never be chosen.

Example: 3 of 5 design variants failing the p95 requirement before weighting even starts is the constraint step doing most of the actual decision work -- the later weighted comparison only has to adjudicate between the two options that were left, which is a much smaller and more honest question than comparing all five.