Lab 15 · The Architect Elevator Pack

The architect's differentiating skill, in the elevator framing this track borrows, is riding between the penthouse and the engine room without changing the story — only the altitude. Anita has made her recommendation on Orbit's copilot platform; now the same decision must survive a board that reads money and risk, an engineering-leadership group that reads trade-offs, and a build team that reads interfaces and definitions of done. AI removes the production cost of writing three artifacts; it does not remove the failure mode that kills packs like this in review: the versions quietly contradicting each other. So this lab makes verification part of the deliverable — you will seed a contradiction, catch it mechanically, fix it, and only then call the pack done.

Time & tier

~60 minutes hands-on · Core tier (maps to module M3). Runs at ~₹0 — design-artifact lab; any chat assistant with a decent context window (free tier or local via Ollama) is sufficient.

Setup

Create lab-15/ with one-pager.md, leadership-deck.md, team-brief.md, and consistency-report.md.

The decision brief, exactly as Anita's working notes state it (all numbers synthetic):

DECISION: Platform for Orbit's customer-facing copilots (orbit-docs-copilot
first, two more copilots on the 18-month roadmap).

OPTION A - BUY: managed AI platform suite (vendor "NimbusAI").
  Vendor-quoted ~Rs 66L/year at projected usage; runs with 2 engineers.
  Live in ~6 weeks. Vendor claim in the proposal: "unlimited evaluation
  runs included at no extra cost." Exit cost: HIGH - prompts, eval sets
  and routing config in proprietary formats; egress rework est. 2 quarters.

OPTION B - BUILD: OSS gateway + agent framework + pgvector on Orbit's
  existing cloud. Est. ~Rs 38L/year infra + the time of 4 engineers for
  2 quarters, then 2 to run. Live in ~14 weeks. Exit cost: LOW -
  every component replaceable; model APIs already abstracted.

RECOMMENDATION (decided): B-hybrid - BUY managed model APIs, BUILD the
orchestration/eval/gateway layer on OSS. Cost path ~Rs 45L/year steady
state. Live in ~12 weeks. Needs 4 engineers for phase 1, 2 ongoing.
KEY RISKS: hiring/retention of the 4; OSS component churn; 12-week
timeline slips if eval tooling is underestimated.
ASK: approve Rs 45L/year run-rate + 2 backfill hires; phase-1 review
gate at week 6.

One instruction before you draft: the NimbusAI "unlimited evaluation runs at no extra cost" line is a vendor claim inside an AI-summarized proposal. Decide during Step 2 what a responsible pack does with it.

Guided steps

Step 1 — Write the decision spine (10 min)

Before any audience-shaped prose, extract the invariants — the facts that must be identical at every altitude. In a scratch section at the top of consistency-report.md, list them: the decision (B-hybrid), steady-state cost (₹45L/yr), timeline (12 weeks, gate at week 6), staffing (4 then 2, +2 hires), top three risks, and the ask. This spine is your single source of truth; every artifact derives from it, and the consistency check will test against it.

✅ Checkpoint: the spine exists with the six invariants; nothing in it is audience-flavored.

Step 2 — The board one-pager (10 min)

Draft one-pager.md with your assistant, then edit hard. Penthouse rules: one page; leads with the business outcome (three customer-facing copilots on one platform) and the ask; at most three numbers (₹45L/yr, 12 weeks, 2 hires); risk stated in exposure terms ("vendor lock-in avoided — exit measured in weeks, not quarters"), not component terms; zero product names except where the board must approve spend on one. And the vendor-claim decision: a board pager that repeats "unlimited evals at no cost" as fact has laundered an unverified vendor claim upward — either omit it or mark it as "vendor-claimed, verification pending." Choose and note your choice in the margin.

✅ Checkpoint: one page, ≤3 numbers, ask in the first or last paragraph, and the NimbusAI claim either absent or explicitly flagged as unverified.

Step 3 — The engineering-leadership deck outline (10 min)

Draft leadership-deck.md as five slides, outline form (title + 3–5 bullets each):

1. 1Contextthe copilot roadmap and why the platform decision is now.
2. 2Options consideredA, B, B-hybrid with the comparison dimensions that mattered (cost, time-to-live, exit cost, team load).
3. 3Decision and rationaleB-hybrid; why exit cost and eval-tooling control outweighed six weeks of speed.
4. 4Risks and mitigationsthe spine's three risks, each with a mitigation and an owner.
5. 5Plan and asks12 weeks, week-6 gate, 4-then-2 staffing, the two hires.

This altitude gets trade-off honesty the board pager compresses: leadership should see what was given up (six weeks, vendor support SLAs) and why that is acceptable.

✅ Checkpoint: five slides, every spine invariant present, and slide 2 shows at least one real disadvantage of the chosen option.

Step 4 — The build-team brief (10 min)

Draft team-brief.md — engine-room altitude. Contents: the component list (gateway, orchestration layer, eval harness, vector store, managed model APIs) with what is decided vs open per component; interfaces and constraints (model APIs behind the gateway abstraction from day one — that is what keeps the exit cost low, so it is non-negotiable); phase-1 definition of done (first copilot serving through the new stack, eval gate in CI, week-6 review evidence); and the same risks translated into engineering terms ("eval tooling underestimated" becomes "eval harness is on the critical path — spike it in week 1"). No business framing; a new joiner should know what to build by Friday.

✅ Checkpoint: every component labeled decided/open, phase-1 DoD stated, and the week-6 gate appears with the evidence it requires.

Step 5 — Plant the inconsistency, then run the checker (15 min)

Now prove your verification step works before trusting it — the same test-the-test discipline as seeding a defect in a pipeline. Deliberately introduce one contradiction between altitudes: change the deck's steady-state cost to ₹52L/yr (leave the pager and brief at ₹45L). Then run this consistency-check prompt with all three artifacts pasted in:

You are auditing three artifacts that describe ONE decision at three
altitudes (board, engineering leadership, build team). Build a table of
every fact that appears in two or more artifacts: decision, costs,
timelines, staffing, risks, asks, scope. Columns: fact | one-pager says |
deck says | brief says | consistent? For each inconsistency, quote the
conflicting phrases verbatim. Distinguish true contradictions from
altitude-appropriate differences in detail (an omission or a rounding at
a higher altitude is NOT a contradiction - flag only conflicts). End
with: CONSISTENT or N CONTRADICTIONS FOUND.

Save the output to consistency-report.md. Confirm the checker caught the planted ₹52L/₹45L conflict with verbatim quotes. If it did not, tighten the prompt (or your artifacts' phrasing) and rerun until it does — a checker that misses a planted defect has no authority over real ones. Then fix the deck back to ₹45L, rerun, and file the clean report below the dirty one.

✅ Checkpoint: consistency-report.md contains both runs — one catching the planted contradiction with quotes, one ending CONSISTENT.

Step 6 — Final pass: altitude integrity (5 min)

Read the three artifacts in order, penthouse to engine room, and check the elevator property: each level adds detail without changing any invariant, and each level answers its audience's first question in its first breath (board: "what do you need and what's the risk?"; leadership: "why this option?"; team: "what do I build?"). Fix anything that reads like the wrong floor — jargon in the pager, business hedging in the brief.

✅ Checkpoint: one sentence written at the bottom of the consistency report: which artifact drifted most from its altitude, and what you changed.

Verification

• Three altitudes, one decisionall spine invariants identical across the artifacts (post-fix); each artifact shaped for its audience.
• Checker validatedthe planted ₹52L contradiction was caught with verbatim quotes before the clean run was accepted.
• Vendor-claim hygienethe "unlimited evals" line appears nowhere as unqualified fact.
• Engine-room usabilitythe team brief's decided/open labels and phase-1 DoD would survive contact with an actual sprint planning session.

Completion artifact

The lab-15/ three-altitude pack: one-pager.md, leadership-deck.md, team-brief.md, and consistency-report.md (spine + dirty run + clean run + altitude note). Self-grading rubric (pass = 8/10):

• Altitude fit (0–3): each artifact answers its audience's first question first; no floor-mixing = 3.
• Spine integrity (0–2): six invariants identical everywhere in the final pack = 2.
• Checker discipline (0–3): planted defect caught with quotes, then fixed, then clean run filed = 3.
• Vendor-claim handling (0–1): NimbusAI claim flagged or excluded, with the choice noted = 1.
• Honest trade-offs (0–1): the chosen option's real disadvantages visible at the leadership altitude = 1.

Client-safe variant

This lab's brief is synthetic, but the elevator pack is exactly the artifact class most likely to contain real client strategy — costs, vendor names, roadmap dates. On engagements: draft with the client-approved tenancy or a local model, never a personal chatbot account; strip vendor pricing and internal cost figures before using any external AI for wordsmithing; and keep the consistency-check step inside the same boundary as the artifacts, since the check requires pasting all three at once — the full pack is the most concentrated leak surface you will create all week. The spine-plus-checker method itself is client-portable and worth showing: it is an audit trail for communication accuracy.

Stuck?

• •
  The assistant's one-pager keeps sprouting technical detail: give it a hard frame — "60-second read for a board member; if a bullet needs an engineer to parse, delete it" — and cut to three numbers yourself.
• •
  The checker flags altitude-appropriate omissions as contradictions: your prompt's distinction clause is being ignored — move it earlier, or add one example of a legitimate omission to the prompt.
• •
  The checker misses the planted defect: paste artifacts in full (not summaries), and make sure the conflicting numbers are both actually present — a summary may have dropped one.
• •
  Five slides won't hold the content: the overflow belongs in the team brief — leadership slides carry decisions and trade-offs, not designs.
• •
  The team brief reads like documentation of the decision instead of instructions: rewrite each section to end with something buildable this sprint or a named open question with an owner.

Answer key (don't read before attempting)

The seeded elements and what they teach:

1. 1The planted ₹52L/₹45L contradiction is a proxy for the real killer. In genuine packs the drift is rarely a typo — it is version skew: the deck gets updated after a leadership review, the pager doesn't, and the board approves a number engineering no longer believes. The defense is structural: one spine file, artifacts derived from it, and a checker run on every revision — not proofreading heroics.
2. 2The checker had to be tested before it was trusted. This is the same discipline as Lab 12's calibration step and Lab 01's seeded defects in the DevOps track: a verification mechanism earns authority by catching a known defect first. If your checker passed the dirty pack, the correct response was to fix the checker, not to assume the pack was fine.
3. 3The "unlimited evaluation runs at no cost" line is this track's signature trap at the communication layer. It is a vendor capability claim, AI-summarized, heading for a board document — the exact path by which hallucinated or marketing-soft claims become organizational commitments. Correct handling: exclude it, or carry it as "vendor-claimed; verify against the contract and rate card before relying on it." Repeating it as fact costs the full rubric point because in real life it costs considerably more.
4. 4The judgment call the checker cannot make. The board pager saying "~₹45L/yr" while the brief says "₹45L/yr steady state after two quarters at higher spend" is not a contradiction — it is altitude-appropriate compression. Deciding which differences are conflicts and which are compression is the architect's edit, which is why the consistency report ends with your note, not the model's verdict.