Introduction

STF (Structured Text Format) is a human-readable format for configuration and data interchange, defined by a normative specification, an exact condition-to-error-code mapping, and a conformance corpus that every implementation runs.

STF

# comments are part of the format
{
  service: `checkout-api`,
  port: 8080,
  enabled: T,
  deploy_after: TIMESTAMP(2026-01-15T10:30:00Z),
  price_cap: DECIMAL(199.00),   # scale is data
  account_id: BIGINT(9007199254740993),
  signing_key: BINARY(SGVsbG8=),
  regions: [`eu-west-1`, `us-east-1`],
}

The problem

JSON has six types, so everything else travels as a string or a binary64 number, and the reader is left to guess from context:

JSON	The question it leaves open
`"created": "2026-01-15"`	A date, or a string that looks like one?
`"id": "9007199254740993"`	A string, or an integer stringified to survive `JSON.parse`?
`"price": 19.99`	Not exactly 19.99 — it is the nearest `binary64`.
`"blob": "SGVsbG8="`	Base64 by convention, decoded by whoever remembers.

Every one of those is answered by a convention held outside the document — a schema, a field naming rule, or a comment in a wiki. STF puts the answer in the document.

STF

{
  created: DATE(2026-01-15),
  id: BIGINT(9007199254740993),
  price: DECIMAL(19.99),
  blob: BINARY(SGVsbG8=),
}

Replacement, not superset

STF deliberately does not inherit JSON's looseness, so some valid JSON has no STF representation: non-identifier keys, a non-object root, NaN and Infinity, and duplicate keys. A conversion that meets one of those fails loudly rather than guessing.

That is the trade. What it buys is that two implementations cannot quietly disagree: keys have one spelling, duplicates are an error rather than a last-one-wins race, every rejection has one documented code, and there is a canonical byte encoding for hashing and signing.

Where to go next

Syntax — documents, keys, strings, numbers, comments, directives.
Data model — the eleven value kinds and how equality is defined.
Typed constructors — DECIMAL, BIGINT, DATE, TIMESTAMP, BINARY.
Canonical form — one byte encoding per value, plus resource limits.
Error codes — the full index.
Record streams — .stfs, for append-only logs.
Command-line tool — check, format, lint, convert, and the language server.
Migrating from JSON — what converts, and what needs a decision first.

Status

The 1.0 draft is complete and every reference implementation passes the corpus. STF is pre-release: the specification may still change incompatibly until 1.0 is tagged.

The normative documents live in the repository and remain the source of truth — specification, error codes, stream profile, and schema.