Exercise 3: Choosing a Structure

Overview

Three artifacts for Lintel, about three hours. This is the first module whose output another person could build from — and it closes the groundwork, so these three documents plus Exercises 1 and 2 should read as one coherent engagement.

Every scoring decision here cites Exercise 2. Keep your ASR list, Utility Tree and NFR budget open; if you find yourself inventing a criterion that is not traceable to one of them, that is a finding about Exercise 2 rather than a licence to proceed.

How to work

In order. The reference architecture is unwritable before the style is chosen, and the roadmap is unwritable before you know what you are sequencing.

Style selection with trade-off table — ~60 min
Reference architecture sketch (C4 Level 2) — ~50 min
12-month roadmap — ~60 min
Self-check — six pass/fail items

Artifact 1 — Style selection with an explicit trade-off table

One table plus one page of reasoning.

Build it in this order, and make the order visible in the document:

Constraints as veto columns. Residency, SOC 2, the personal data in tickets, the existing estate. An option that violates one is eliminated, not scored low.
Weighted criteria from your ASRs, with the weights taken from your Utility Tree rather than invented here.
At least four candidate options, including the one that is tempting for a bad reason and one you expect to reject.
A capacity column. What each option demands in engineering time, and what that displaces.
The decision, with the runner-up named and the condition under which it would have won.

Then run change-case analysis against your chosen structure: three changes you expect within a year, and how many boundaries each crosses. If two of them cross three boundaries each, revisit the boundaries before you draw anything.

The test: a reader should be able to tell which option you wanted before you started, and see that the criteria were fixed anyway.

Artifact 2 — Reference architecture sketch

One C4 Level 2 diagram plus two short lists.

The diagram shows containers — where the AI components sit relative to the Rails monolith and the Go services, with the external dependencies that matter. Not internals. If a reader can tell how retrieval works inside, you have dropped a level.

Then the two lists that make it a reference architecture rather than a design:

Rules that hold without exception, each with its reason. Aim for four to six. Anything you cannot justify in a sentence comes out.
What is deliberately left open — the decisions the teams keep. If this list is short, you have written an implementation.

Include the degradation ladder as components, not as error handling. Rungs below the first are structural elements with their own entry conditions, and the bottom rung must contain no model.

Write the per-second rate somewhere in the document. It is one line, and it justifies most of what you chose not to build.

Artifact 3 — 12-month roadmap

One page. Four quarters, sequenced against the fact that capacity is already fully consumed.

Required content:

Each move stated as a bet — the bet, why now, what you expect, what would falsify it, cost to reverse.
Ordering by reversal cost, with any violation of that principle explained. (There will be at least one, and the residency decision is the likely culprit.)
A thin vertical slice first, not a complete horizontal layer.
One displacement statement per quarter: to do X by D, Y slips to E, owner named, agreed yes/no.
What each stakeholder gave to arrive at this sequence — at minimum for the product, compliance and engineering leads.
Review dates for anything deferred.

The test: hand this to the VP Engineering and he can tell you whether it is possible. A roadmap he cannot evaluate is not a roadmap.

Self-check

Pass or fail.

The style choice cites specific ASRs from Exercise 2, not general preference. Named ASRs, in the table.
The trade-off table includes at least one option you rejected and why. Real, not a straw man, with the criterion that killed it named.
The reference architecture respects the existing monolith rather than assuming it away. It appears in the diagram, reached through an interface.
The roadmap accounts for zero spare capacity and says what gets displaced. With an owner’s name on each displacement.
At least two quality tactics are named as deliberate choices with their costs stated. In the tactic, cost, requirement form.
The design process is written down as steps you could hand to someone else. Including your triage rule for when it applies.

If something failed

If the trade-off table has no real rejected option, you probably decided first and documented second. The fix is not cosmetic: pick the option you dismissed in your head without writing down, give it a fair row, and see whether your conclusion survives. Sometimes it will not, and that is the exercise working.

If the roadmap fits comfortably, you have not checked it against capacity. Every engineer at this company is committed; a plan that needs nobody to stop doing something is a plan that assumes capacity that does not exist.

If the reference architecture reads like a design, the “deliberately left open” list is too short. Go back and move at least three decisions from your diagram into the teams’ hands — and notice how uncomfortable that is, because that discomfort is the Exercise 1 charter question arriving in concrete form.

→ A structure and a sequence exist. Next, the AI components inside them. — Exercise 4, after reading Handbook Module 4.