When Buyers Ask for Eval Packs Before the POC

Buyers in Latin America are asking for dated eval evidence before they expand a POC—not after. That is not bureaucracy. It is risk management under thin margins and high switching costs.

What changed

Coastal sales motion assumed demo first, proof later. Mid-market and enterprise buyers in Brazil and the broader LatAm corridor increasingly ask for:

• A golden set in the languages they serve
• Latency and cost numbers on that set
• A clear data-path story (where prompts, logs, and embeddings live)

If your pack is vibes and screenshots, you are late.

The one-page eval pack

Keep it boring and dated:

1. Workload — the one workflow under test (support triage, document Q&A, etc.)
2. Golden set size and languages — e.g. 25 prompts, PT-BR + ES
3. Pass/fail definition — correctness, refusal, tool-contract shape
4. Cost and latency — p50/p95 and token spend under a stated FX assumption
5. Data path — one diagram: client → model/vendor → storage → retention

Engineering owns the numbers. Sales owns the calendar. Same artifact.

Why this compounds

Eval packs turn Eval Gates Before You Ship into a commercial asset. They also force the uncomfortable truth early: if you cannot measure Portuguese quality, you cannot promise it.

Janeiro’s stance stays the same—evidence before expansion. Calm, dated, reusable.