OBDS is an open, implementation-ready specification for determining which governed brand truth is authoritative and applicable to a specific AI task, and whether execution may proceed.
Which truth applies, and may this run?
Brand truth is not one static document. It varies by market, channel, audience and time, some of it is approved and some of it is not, and parts of it are simply not known. OBDS is the layer that decides which part is authoritative for the task in front of it.
What problem it solves
An AI system asked to write a product claim can retrieve a perfectly relevant paragraph and still be wrong: the paragraph may apply to a different market, may have expired, may be contradicted by an approved rule, or may rest on a fact the brand has never approved. Retrieval scores relevance. It does not decide authority, applicability or permission. OBDS does.
How it differs from retrieval
| Layer | Question it answers |
|---|---|
| RAG / retrieval | What information looks relevant to this request? |
| OBDS | Which governed information is authoritative and applicable to this task, and may execution proceed? |
Knowledge access is not governed applicability.
OBDS does not replace retrieval. It determines which retrieved or declared brand truth is authorised and applicable to the current task. The two compose: retrieval narrows the field, OBDS decides what is allowed to reach the model, and what must stop the task instead.
A concrete example
A brand manifest declares two elements. structure.brand is a fact in state defined. context.efficacy-claim is in state unknown: the brand expects a value there and does not have one. A build target requires both.
status failed
requirements structure.brand defined pass
context.efficacy-claim unknown fail
error OBDS-BUILD-REQUIRED-NOT-DEFINED
context files none written
model calls 0Real output of the published fail-closed example, exit code 2.
The unknown element is never guessed, never widened to a neighbouring scope and never quietly dropped so the build can succeed. No Compiled Brand Context exists, so there is nothing to assemble a model input from, so no model is called. That is what fail closed means here: the absence of required truth stops the process before generation, rather than producing fluent output from an assumption nobody approved.
The vocabulary
- Governed brand truth — brand information a named human role has approved, carried with its provenance.
- Authority — which declaration wins when more than one could apply.
- Applicability — whether a declaration applies to this scope and at this time.
- Scope — the declared dimensions a truth is valid in: locale, market, channel, output type and six more.
- Required truth — what a target declares it cannot proceed without.
- Fail closed — missing required truth stops execution instead of being filled in.
- Compiled Brand Context — the immutable artefact for one task, hashed and reproducible.
Limits
OBDS begins at an approved Brand Manifest and ends at a decision record. It cannot tell whether an approved value is true in the world, and it cannot detect a build target that under-declares what it should have required: both build successfully with valid hashes. It is brand-specific and is not a general enterprise policy engine, a prompt format, a RAG system or an AI operating system. It does not render anything; it declares the permissible design space that renderers create inside. The published limits document states this in full.