Shared vocabulary for split-context verification: nodes, webs, rotation, fabrication, and the failure modes each of them has. Where this page and a description elsewhere disagree, this page is the one that governs.
A verification mechanism requiring two or more logically independent contexts held at the same time in an agent's processing environment, where no single context holds enough to produce the output, the output is determined by the relationships between contexts, the meaning of information in any one context is established by others in the web, and adding, removing or modifying any context changes the output.
A single independent source of information in the web. Each node is authored separately, stored separately, and modifiable independently. A node can be a prompt, a document, a database record, a skill file, a boot sequence, or a policy clause that matters downstream.
The structure of relationships between nodes. Two nodes make a single edge. Three can form a chain or a triangle. As the count rises the topology gets richer and the verification gets harder to compromise.
What the agent produces to show it processed the web. Not stored in any single node; produced by processing several at once.
The output changes if any node is modified, added or removed. A stale copy of one node, or a missing node, produces a different answer.
Defines the test: what to look for, how to combine it, what counts as the output. It does not contain the answer.
Contains information needed to produce the output, without identifying which of its contents matter. Significance is established by instruction nodes, or by relationships between content nodes.
Functions as instruction and content together. A document saying "your standard is WCAG 2.2 AA" and also "always verify your standard against the current published version" is doing both jobs, and hybrid nodes make denser topologies.
An ordered set of nodes loaded in sequence during boot or execution. The web in its temporal form.
Leaking one node does not reveal the output, because the output depends on relationships the leaked node does not describe. The property weakens sharply when the answer appears verbatim in a single content node.
A plausible but incorrect output is caught because the multi-node structure makes fabrication detectable. Richer topology makes it harder to fake.
Changing any node invalidates the current output. In an N-node web there are N axes of rotation, each independent of the others.
Nodes carry both verification-significant information and governing directives, so running verification loads the directives alongside the content. A pass indicates the expected output matched. It does not guarantee downstream obedience to every behavioural rule.
Information inside a node that acts as an instruction, but only once the agent has loaded other specific nodes. Without them it reads as ordinary content.
Because the output depends on every node, a stale version of any one produces an incorrect answer, so a single check confirms all N documents are current.
The agent lacks content nodes and fabricates from training data. Detected because the fabricated output does not match the expected value.
The agent passes at one level and fabricates at a lower level where no web exists. Mitigated by layering webs at every level.
An adversary obtains every node and can produce the output. Mitigated by separate storage, independent access controls, and frequent rotation.
An adversary obtains some nodes but not all, and cannot produce the output if the missing nodes matter to the configured relationship. Resilience depends on which nodes leaked, not only how many.
The verification information already exists in training data and the agent may guess correctly. Mitigated by using novel, invented information.
The web needs more tokens than the agent can hold, so it cannot process every node. Mitigated by designing within the window.
An adversary injects a fake instruction node mid-session. Mitigated by delivering instruction nodes over injection-resistant channels.
A content node loads incompletely and the output is wrong. Detected by the check failing.
The agent produces the correct answer in an unexpected format. Mitigated by normalising output before comparison.
An adversary infers the structure from agent behaviour. Low severity: knowing the shape without the node contents does not let anyone produce the output.
Every required node fits in the window, but the model under-uses middle or low-salience content, the effect known as lost in the middle. Verification can still fail when the expected answer requires combining exactly those pieces.