A source of truth is the authoritative source for a particular piece of information. In an agent system, authority depends on who owns the value and where it came from—not how often it appears in context or whether a generated copy has the right format. section 4.2 section 11.2.4
Start with authoritative sources the work already owns. For a coding agent, the repository, rules files, dependency graph, version history, and developer preferences may preserve much of the context needed for later tasks. Add a dedicated memory layer when extraction, isolation, updates, recall, or cross-system reuse justify another control surface. section 4.1.3
When information has multiple copies or retention tiers, define which copy is authoritative and what promotes, supersedes, invalidates, or archives it. Separate candidate memories from active instructions: a frequently retrieved learned preference can remain subordinate to a human-owned rule. Track stale reads, harmful promotions, and decisions that override a human-owned source. section 4.2
A valid argument with an uncertain origin
A receipt contained an exact template identifier required by an import service. After the receipt entered conversation history as prose, a later model call regenerated the import arguments. The resulting call could pass JSON Schema validation without proving that the identifier came from the receipt. A well-formed value aimed at an unintended entity was the more serious risk; retries or prompt repair would not establish ownership or provenance. section 11.2.4
matching receipt with exact template_id -> prose in conversation history -> model regenerates import arguments -> schema validates argument shape -> import receives a value with unproven provenance
Follow the identifier from the receipt through conversation history to an import call whose argument shape is valid but whose provenance is unproven. section 11.2.4
At task boundaries, identify the principal, authoritative inputs, environment, permitted effects, acceptance owner, and recovery rule. This makes the inputs and authority boundary explicit before the agent acts. section 10.1.1












