2026-11-18datamarket-dataproduct

Every number with a receipt

DecisionGateway's market-data catalog ships numbers only when they arrive with the exact sentence they were read from — English, Russian, Arabic, German, Spanish. Here is what that buys you, and what it costs us.

Most market numbers on the internet arrive as naked digits. No source, no sentence, no name willing to stand behind them. If you have ever tried to trace one back to what was actually said, you know the feeling.

We do it differently, because our whole graph is built on attribution. When a public wire item — a regulator statement, a company release, a market note — contains a market number, it is only kept together with the verbatim sentence it came from. If we cannot quote the sentence, the number does not ship.

What the catalog shows

The public market-data catalog at dgw.tmarket.io/market-datums carries the verified subset of those numbers:

  • The figure, first. Each row leads with the parsed value, its unit and currency, and the instrument it refers to.
  • The receipt. Below the figure, the verbatim source sentence — the exact span the number was read from, in the language it was published in (observed so far: English, Russian, Arabic, German and Spanish).
  • The instrument. An instruments directory resolves names to stable identities. The asset-resolution backfill is still rolling out — most rows are not yet tied to a resolved instrument, so this is a direction, not a finished surface today.
  • Live counts. The catalog header carries a live count rather than a stale marketing figure — as of 2026-11-18 there are 19 figures on record, and the count grows as more numbers pass verification. (llms.txt is a static index of the surface, not a live ledger.)

The API is open and spec-first: GET /v1/market-datums, /v1/market-datums/{id}, /v1/market/instruments and /v1/market/facets are published in the public OpenAPI document under a deny-by-default allowlist. No key required for reads.

Why so few numbers, on purpose

A catalog of market figures is worthless if a meaningful share of its rows are wrong. Every datum passes deterministic invariants before it is stored: the quote must contain the raw tokens verbatim, numbers are re-parsed from the text, direction must match sign, ranges must be internally consistent, and duplicates inside one post collapse. Numbers that fail never appear here — not as "low confidence" rows, not as asterisks. That is the trade: fewer figures, each one defensible.

Why it matters for us

The same discipline — who said it, when, in what words — is what powers the rest of the graph: claims, actors, decisions, forecasts. The market catalog is the same contract applied to the market numbers the wires already publish.

This is not a stylistic preference. Independent evaluation of AI models shows verbatim quoting is exactly where they fail: a 2026 study of clinical QA found most models attach verbatim quotes to over 90% of their claims, yet full substantiation lags far behind — Claude Opus-class models quoted verbatim for 98.0% of claims while fully substantiating only 37.1% — with failures splitting into "the quote wasn't verbatim" and "the quote doesn't support the claim" (arXiv:2609.15964). The fix at the source is to never let the pair break apart: store the number together with its sentence, or don't store it.

If you want machine-readable access, the four endpoints above are the whole surface. If you want the reasoning behind a number, follow any row to its source sentence — that is the product in one line.

Methodology note: catalog counts are live (the catalog header; llms.txt is a static index); the 19-figure figure was current on 2026-11-18 and will drift upward. Sources are public wires and releases; every figure links back to its verbatim sentence.