Translating Across Audiences

Overview

The same architectural decision has to reach people who care about entirely different things. The common mistake is to write one document and adjust its tone.

Four framings of one decision differ in substance, not register. Each answers a different question, and most of the design is absent from each.

Four audiences, four questions

The reframe that makes this tractable: you are not explaining your decision. You are answering their question.

Audience	Their actual question	What the framing leads with
Whoever holds the budget	What does it cost, what comes back, what is the risk to the money?	Unit economics, the ceiling, and what happens if volume doubles
Whoever owns the engineers	What capacity does this need, what does it displace, who maintains it?	The displacement statement and the ongoing maintenance load
Whoever can block on compliance	What data moves where, what is the blast radius, can I evidence it?	The data flow, the enforcement points, and the audit evidence
The team who will use it	What changes for me on Monday?	The workflow, the handoff, and what they are not expected to absorb

Note what is not on that list: the architecture. None of these four audiences is asking what the system looks like, and leading with a diagram answers a question nobody asked.

Omission is the skill

graph TD
  A["<b>The decision</b><br/>full reasoning,<br/>all trade-offs,<br/>every constraint"] --> B["Budget framing<br/><i>~15% of it</i>"]
  A --> C["Capacity framing<br/><i>~15% of it</i>"]
  A --> D["Compliance framing<br/><i>~15% of it</i>"]
  A --> E["User framing<br/><i>~10% of it</i>"]

  classDef whole fill:#5B8FB9,stroke:#4A7A9E,stroke-width:2px,color:#ffffff
  classDef part fill:#7D9A4E,stroke:#5F7A38,stroke-width:2px,color:#ffffff
  class A whole
  class B,C,D,E part

Four framings, each discarding most of the decision. The discarding is the work — and it feels like dishonesty the first several times you do it.

Each framing leaves out the trade-offs that audience does not own, the alternatives they would not evaluate, and the mechanisms they will never touch. That is not simplification — it is selection, and it is the part that takes judgement.

Two tests for a framing:

Could the recipient act on it? If the document ends and they do not know what they are being asked to agree to, fund or approve, it has failed regardless of accuracy.

Would they recognise their own concern in it? The first paragraph should contain the thing they worry about. If it opens with your concern, they will read the rest as yours.

Why detail-dumping reads as low confidence

Worth stating plainly, because the instinct runs the other way.

A forty-page document sent to someone who asked a one-page question carries an implicit message: I could not tell what mattered. It transfers the work of selection to the reader, and the reader concludes — usually correctly — that the author had not finished thinking.

Confidence is demonstrated by what you leave out. An architect who can answer the budget question in one page, with the arithmetic and the risk, has shown mastery more convincingly than one who attaches the full design.

The related error is defensive completeness: including every caveat so nothing can be held against you later. It reads as hedging, and Module 1 warned that hedging stacks read as weakness rather than rigour.

The one thing that must not vary

The decision and its reasoning are identical across all four framings. Four presentations of one decision — never four different decisions.

This is where the consistency mechanism from the first topic becomes load-bearing. These people talk to each other. If the budget holder hears that cost is the deciding factor and the compliance owner hears that residency is, the first time they compare notes you have lost both of them, and the loss is not recoverable by explanation.

A practical safeguard: write the single-sentence version of the decision and its primary reason first, then write the four framings around that same sentence. If a framing requires changing the sentence, you have discovered either a framing that is wrong or a decision that is not yet settled.

One decision, four framings

The Module 4 decision: hosted inference, pinned in-region, behind an interface, with the provider selected against documented criteria.

The sentence that does not change: We will use hosted inference inside the region the data originates in, because the enterprise contracts forbid egress and we have no capability to operate accelerators ourselves.

For Dev Raghunathan — cost, return, risk to the money. Leads with: cost per deflected interaction of about $1.54 at p50 against a $2.50 p95 ceiling, inside a monthly inference budget of $8,000. Includes the finding that the company-wide $3.00 target does not close on deflection alone, and where the remaining value has to come from. Includes the hard spend cap and the degradation behaviour. Omits: chunking, the agent’s pattern, the handoff protocol.

For Tomas Beck — capacity, displacement, maintenance. Leads with: this needs one engineering slot in the first quarter, and the billing read-path extraction slips from Q1 to Q3 as the named displacement. Then: self-hosting was rejected because it would require GPU operations capability his teams do not have and cannot acquire alongside the migration. Then the ongoing maintenance load. Omits: unit economics, the eval methodology, the compliance mechanism.

For Lena Fischer — data flow, blast radius, evidence. Leads with: the three gates. Inference pinned in-region with contractual guarantee; retention terms assessed against SOC 2 and GDPR; and per-request recording of where processing occurred, retained for the audit cycle. Then the erasure table reaching embeddings, caches and traces. Then the enforcement points, marked deterministic or probabilistic. Omits: cost, capacity, the roadmap.

For Grace Adeyemi — what changes on Monday. Leads with: her agents receive handoffs containing the conversation, the sources, what the system concluded and what it already did. Then: every credit comes to them as a confirmation request, and here is the expected volume. Then what they are not expected to absorb. Omits: essentially the entire architecture.

Where they would compare notes, and what holds. Dev hears that residency constrains the model choice and therefore the cost; Lena hears that the cost ceiling constrains how much checking is affordable. Those are the same fact seen from two sides, and they reconcile — which they would not if one of them had been told cost was the deciding factor and the other that compliance was.

Feeds the workbook

This is the three one-page framings artifact directly, and the rubric is explicit: they must differ in substance, not just tone. The test is whether each one omits something the others lead with.

→ Four audiences, handled. One person remains, and they have seen something like this fail before. — Earning trust from the sceptic