Lesson 19.2 gave you five questions and a source card. This lesson attacks the mistake that empties the card fastest: picking a source *because it is free* and skipping every other question.

Free tells you the entry price. It tells you nothing about whether the data covers your question, arrives on time, stays up, covers history, permits your use, or survives the provider's next redesign. Beginners optimize for price. Builders optimize for fit — then check price.

Five labels people confuse with "free"

• No-cost public access — the owner publishes it openly (e.g., SEC EDGAR: public, authoritative, read-only filing data). Still has rules: fair-access guidance, no browser CORS on data.sec.gov, filings — not prices.
• Free tier — a commercial product lets you try a bounded amount. Beyond the bound: throttled, cut off, or billed. The bound *is* the contract.
• Trial credits — temporary money that expires. Prototypes built on expiring credits die on schedule unless you plan the exit.
• Open data — published under an open license with reuse rights (often attribution, sometimes share-alike). Open still has terms; read them.
• Unofficial endpoint — someone else's scraper, a leaked key, an undocumented consumer-site feed. No contract, no changelog, no one to call. This is not a tier. It is a gamble.

Two sources can both cost "$0 today" and be completely different strategies. The card from 19.2 exists to force that distinction into writing.

The four hidden costs (the iceberg)

            ～～～ waterline ～～～
            Visible: $0 / free tier
   ─────────────────────────────────────
   │ 1. LIMITS      │ rate caps, quotas,   │
   │               │ key requirements,     │
   │               │ fair-access throttles │
   │ 2. DOWNTIME   │ outages, deprecations,│
   │  + HISTORY    │ shallow backfill,     │
   │               │ no vintage/revisions  │
   │ 3. RIGHTS     │ attribution,          │
   │               │ no redistribution,    │
   │               │ no commercial use     │
   │ 4. MAINTENANCE│ silent schema changes,│
   │               │ broken scrapers,      │
   │               │ migration on rename   │
   ─────────────────────────────────────
   What sinks prototypes lives below the line.

In plain language:

1. Rate limits and uptime. A free tier that allows 10 requests a minute is fine for a demo and fatal for an hourly alert product. Open-Meteo is generous but still rate-limited; design the call schedule before you promise freshness. 2. Downtime and missing history. Free endpoints are often the first throttled and the last restored. Shallow history ("last 30 days only") quietly kills trend features. 3. Attribution and redistribution restrictions. "Free to view" is not "free to store, republish, or sell." Some licenses require attribution on every display; some forbid redistribution entirely. 4. Maintenance when the provider changes. Unofficial feeds change without notice. Official APIs version and announce. Maintenance is a cost measured in your evenings.

A concrete contrast

Take one product fact — "tell me something about company X's financial position":

EDGAR is authoritative for *what the company disclosed* and useless for *what the stock did at 10:32 a.m.* FRED is superb for *economic context* and useless without its metadata. The unofficial endpoint looks easiest and carries every hidden cost simultaneously.

This is also why a small paid provider can be cheaper than engineering around bad data. A documented feed with history, uptime, support, and a stable schema can cost less than the weeks you spend caching, re-scraping, backfilling, and apologizing. That is arithmetic, not a sales pitch: this course never tells you to buy anything. It tells you to *compare total cost honestly* — and the comparison sometimes favors paying.

Worked prototype decision — with numbers

Product fact: "7-day high/low for 3 trip cities, refreshed hourly, plus a 30-day 'unusually warm?' note." Candidates: Open-Meteo vs. a scraped search-result snippet.

1. Rate-limit math (write it down). Need: 3 cities × 24 refreshes/day = 72 calls/day, ~2,160/month. Candidate A docs (check at build time — numbers below are the worked example, re-verify): free use allows ~10,000 calls/day with fair-use throttling. Headroom: 72 / 10,000 = 0.7% of quota — passes with margin for retries. Schedule written on card: cache 1 hour, burst max 1 req/10s. Candidate B: terms allow 0 automated calls (presentation page, no automation right). Need 72 > allowed 0 — fails before any code. That is the whole decision in one fraction.

2. History depth (what breaks). Need: 30 days of history for the trend note + 7-day forecast forward. Candidate A: forecast API (7–16 days) + archive API (years back) — covers both with latitude, longitude, timezone, temperature_2m_max fields kept per row. Trend feature lives. Candidate B: snippet shows today + 6 days, no history endpoint, no timezone field — trend feature dies on day one. Hidden cost #2 (downtime + history).

3. License line (copy the sentence). Candidate A card line: License: Open-Meteo free for non-commercial use with attribution — "Weather data by Open-Meteo.com" on display; commercial / high-volume terms checked 2026-09-11 at https://open-meteo.com/en/docs (re-check at build — this course teaches the method, not a stale table). Candidate B card line: License: unknown — search page ToS forbids scraping; treat as prohibited until cleared. No clever header rotation fixes a "no" in the terms.

Decision + exit path (written out, non-price reason):

Cover the price row: A still wins on coverage, freshness, license, and history. That is how you know the reason is real.

The safe starter pattern

You do not need the final enterprise contract for a prototype. You need a bounded experiment:

1. Bounded — one product question, one source, read-only calls. 2. Logged — every response stored with source URL, owner, retrieval time, and fact date/period. 3. Limited — a usage cap and schedule written down (requests per day, cache duration). 4. Exitable — a named fallback or migration path if the tier ends, the schema changes, or the terms tighten.

Prototype read-only. Log provenance from day one (Class 18's rule: URL, owner, date/period, unit, retrieval time). Set the limit *before* the first call. Keep the exit open. That is a strategy; "$0" is not.

Browse candidates in the ByeBuy Data directory, then evaluate them with 19.2's card — the directory discovers, the card decides.

Practical exercise

Compare two possible sources for one product fact (e.g., inflation context via FRED vs. a scraped news page; weather via Open-Meteo vs. a search-result snippet).

1. Fill a two-column scorecard (19.2) for both. 2. Choose one for a prototype. 3. Write the reason as a non-price sentence: coverage, freshness, license, history, or reliability — never "it is free."

Finish line: two mini source cards + one chosen source + one non-price reason in writing.

Verify: cover the price row of both cards. Is the winner still the winner? If the choice collapses without "$0," re-decide.

Common failure mode: choosing the unofficial endpoint "for now, we will switch later." Later arrives as an outage. If you must start unofficial, write the exit date and fallback on the card today.

Check your understanding

1. Name the five labels people confuse with "free" — and which one has no contract at all. 2. Your prototype needs five years of history; the free tier offers 90 days. Which hidden cost is this, and what breaks? 3. What are the four parts of the safe starter pattern?

Transition

You can now judge one source and see past its price. Real products need several sources at once — filings plus macro context plus user criteria plus your own notes. The next lesson turns scattered cards into a single source map your whole project (and your AI agent) can follow.