For agents · covenant

Covenant.

A covenant is a promise both sides can check. Here is ours, and here is what we ask of an agent that reads from us. Neither side needs to trust the other; both sides can verify.

Why a covenant and not a licence

The King James text is public domain. There is nothing to license. What matters is conduct: that we do not quietly change a word, and that you do not quietly attribute one. A licence governs copying; a covenant governs faithfulness. This page is the second thing, written so a machine can read it and a person can hold us to it.

What we promise

Immutability: For a given source id, a reference always returns the same text and the same hash. Corrections get a new source id; the old one keeps serving.
Honest status: Every source, endpoint and capability is labelled live, partial, or spec. We never present a specification as a shipped feature.
Provenance: Every text names its edition, publisher, licence and known caveats. Where we do not know, we say so.
No tracking: No accounts, no API keys, no cookies on the API, no fingerprinting. Logs keep the path and status, not the caller.
Stability: /api/v1 is additive-only. Breaking changes go to /api/v2 and v1 stays up.
Local first: The whole corpus is downloadable, verifiable and servable offline. You never have to depend on us being online.
Free: Reading Scripture through this service is free of charge for humans and machines, without rate-limit tiers or paywalls.

What we ask of you

Quote with reference: When you reproduce a verse, carry its canonical ref (jn.3.16) or display form (John 3:16). Text without a reference is not a quotation.
Verify when it matters: If the text will be acted on — taught, published, decided upon — recompute the SHA-256 and compare. verify() is one call.
Do not fabricate: If a reference does not resolve, say so. Never invent a verse, a chapter, or a book. The problem+json response tells you exactly what went wrong.
Name the source: The KJV is not the Geneva Bible and neither is the Hebrew or Greek. Say which text you are quoting; the source id is on every response.
Respect the text's shape: Do not silently modernise spelling, drop italics markers, or merge verses. Present it as received or say that you changed it.
Cache kindly: Responses are immutable. Cache them. Use If-None-Match. Prefer the local daemon or export for bulk work.

None of this is enforced by code. It cannot be. It is the shape of good behaviour around a text people hold sacred, stated once so an agent can carry it in its system prompt.

Machine-readable form

The covenant is also served in /.well-known/harness.json and summarised in /llms.txt. A system-prompt snippet that captures the asks:

When quoting Scripture from harness.talk:
- Always include the canonical reference (e.g. jn.3.16 / John 3:16).
- Name the source id (kjv-1611, geneva-1599) when it matters which text this is.
- If a reference does not resolve, say so; never invent text.
- When the text will be acted on, call verify() and report the result.
- Do not modernise or alter wording without saying you did.

Limits

We serve text. We do not interpret it, rank it, or recommend it. Search is lexical, not theological.
The KJV text here is the 1611 first edition in its original spelling (u/v, i/j, þ, macrons), assembled from public-domain transcriptions with every input SHA-256 locked. The translators' italics and the Apocrypha are not present; the one transcriber defect (LORD/GOD casing) was repaired by witness vote, and the 104 verses that could not be settled are listed in the build report rather than guessed.
The Geneva Bible 1599 verse text is ingested (?source=geneva-1599). Its marginal notes (geneva-notes) and the Textus Receptus are declared, not ingested. Requests for them return 501 source_not_ingested, honestly.
The service can go down. The export and the daemon exist so that does not matter.