Three parts, and a check that runs before any user-facing work. The instruction says what to find, the content holds it, and the expected output exists only where they intersect.
The verification output is not stored in any single node. It is produced by the agent processing several nodes at once, and it changes when any one of them changes.
The instruction context. A system prompt, a rule set, a versioned directive loaded from a database. It tells the model what to find, or how to interpret what it has. An instruction node does not contain the answer. It contains the question, the formula, or the selection criterion. If it fails to load, the model proceeds without it and does not report the absence.
The content context. A document, a knowledge file, a memory retrieved from storage. It holds the raw material but does not identify which of its contents are verification-significant. Significance is established by the instruction node, or by the relationship between several content nodes.
The network where lock and key intersect, extended to any number of nodes. The expected output should match only when every required node was staged. A mismatch suggests broken web integrity even when the answer reads as plausible.
Each independently sourced: a database, a file store, an API, a user upload. Nodes may be instruction, content, or hybrid. Every node participates in web integrity, so changing any one of them changes the expected output when the relationships are held constant.
The nodes arrive in the context window, subject to capacity. One or more instruction-bearing nodes describe what to find, how to combine it, or how to traverse a document stack. The model has to resolve references between nodes to answer, because the significance of text in one node is established by another node rather than by any label inside it.
A string or structured value, defined by the web topology and the current versions of every node.
A match indicates the required nodes were present and their joint relationships were interpreted closely enough to land on the expected value. A mismatch indicates a missing node, a stale node, or failed cross-node comprehension. The system can halt, retry, alert, or degrade according to policy.
A missing node produces plausible but incorrect output. A stale node produces output consistent with an older configuration, which is the harder case because the answer is well-formed and drawn from real material. Partial comprehension produces surface-consistent text without the correct cross-reference. Fabrication produces confident output from training priors when the full web was never processed.
The Karl incident on 15 March 2026 was the first of these caught in production. The agent had the document listed in project knowledge and could recite related principles, but never loaded the document itself, and when asked for a word that appeared only inside it, produced a plausible invention with full confidence.
The honest version of the guarantee is narrow and worth stating plainly, because a verification mechanism that oversells itself is worse than none: a reader who believes a pass means more than it does will stop looking.
Continue to the core properties, which set out what the structure buys and where each property stops.