What the FMP Plan Actually Serves

Before any design leans on an endpoint, we call it once with the real key and record what comes back. The first pass ran on 2026-08-28: roughly 45 billable requests, including two binary searches, and it permanently retired five open questions that three documents had been carrying as assumptions. Two more passes followed on 2026-08-29 (forex), 2026-08-30 (ETFs) and 2026-08-31 (bars, delisted names, constituents).

The reason the ritual exists is that a design document built on a wrong belief is more expensive than 45 requests. One of those beliefs — that the news archive reached 2019 — turned out to be off by six years. The archiver is what that correction cost.

The price endpoint is the deep one, and its depth depends on how you ask. With no from or to the default window is exactly 1,254 rows — five years — and the first probe read that as the plan's limit. Pass an explicit from and the same endpoint returns SPY from 1993-01-29, AAPL from 1990, EURUSD from 1999-01-01. Prices are split-adjusted: AAPL in January 1990 comes back at $0.33.

The dividend-adjusted endpoint has a different ceiling: 5,000 rows per response. That is the exact defect Gate 2 caught on its first attempt — the benchmark equity curve was truncated and the comparison failed. The fix was to page, not to widen the tolerance.

Delisted names are the other trap. Lehman, Enron, Yahoo, WaMu and pre-bankruptcy GM all return 200 with an empty array. A consumer that cannot tell "no data" from "broken" will backtest straight through a hole in its history, which is why "empty and error must be distinguishable" became a written rule in the GMI contract, and why survivorship remains our problem, not the vendor's.

Line the two columns up and a rule falls out. Everything that returns 200 is a public fact that is cheap to store: a closing price, a dividend, a split, a headline and its snippet, an ETF's stated sector weights. Everything that returns 402 is copyrighted, licensed from a third party, or expensive to maintain: press releases and transcripts are someone else's text, index membership is a licensed product, one-minute bars are a storage bill, and full ETF holdings are issuer data that changes daily.

This is our reading of the pattern, not a vendor statement — but it has predicted every probe since it was written, and it is now the first question asked before an endpoint is designed against: is this a fact, or is it someone's product?

The cost model books FMP at "$0 marginal if shared with GMI's existing plan." The probe split that in two. For news it holds: stock, general and grade news all return 200 on the existing key. For transcripts it fails: the demand/supply spec books them at $0 and they are 402, so backlog and RPO extraction have no source on this plan. Decision 4e in the implementation plan is still open; FMP support was emailed 2026-09-02.

And the third thing the assumption never addressed, because nobody had looked: the news archive is one rolling year. from=2025-08-29 — 364 days back — returned 200; from=2025-08-28 returned 402. Paging without a date range bottoms out at 92 days. That is not a pricing problem, it is a design problem, and it is why the archiver was built first.

News yes, press releases no, transcripts no, 1-minute bars no, news history capped at a rolling year. That pattern is consistent with the vendor's Starter tier. Whatever the invoice says, this is the capability envelope the design is built inside.

• Archived the feed the same week — 364 days, 627,754 items, before the window ate another month
• Always pass from/to, with a hard floor at today − 364, because the boundary errors rather than terminating
• Did not buy Ultimate reflexively — price the gaps against Tiingo, Polygon and EDGAR first
• Re-run the ritual before any phase that depends on an endpoint; tiers change under you