Glossary of Terms

Shared vocabulary for split-context verification: context nodes, webs, rotations, fabrication, and verification outputs. Hover terms on any page mirror these definitions.

Lock Key Verify Failure
Core Concepts

PassWeb

A verification mechanism requiring two or more logically independent contexts held simultaneously in an AI agent's processing environment, where: (a) no single context contains sufficient information to produce the verification output, (b) the verification output is determined by the relationships between contexts, (c) the meaning of the information in any one context is established by one or more other contexts in the web, and (d) either the addition, removal, or modification of any context in the web changes the verification output.

Core Concepts

Context Node

A single independent source of information in the PassWeb. Each node is authored separately, stored separately, and modifiable independently. A context node can be a prompt, a document, a database record, a skill file, a boot sequence, a policy clause that matters downstream, or any retrievable unit of information.

Core Concepts

Web Topology

The structure of relationships between context nodes. In a two-node PassWeb, the topology is a single edge. In a three-node PassWeb, the topology can form a chain or a triangle. As the number of nodes increases, the web topology becomes richer and the verification becomes harder to compromise.

Core Concepts

Verification Output

The information the agent produces to prove it processed the web. The verification output is not stored in any single context node. It is produced by the agent's simultaneous processing of multiple nodes.

Core Concepts

Web Integrity

The property that the verification output changes if any node in the web is modified, added, or removed. A stale copy of any single node, or a missing node, produces a different output.

Node Types

Instruction Node

A context node that defines the test: what to look for, how to combine it, what constitutes the verification output. An instruction node does not contain the answer. It contains the question, the formula, or the selection criterion.

Node Types

Content Node

A context node that contains information needed to produce the verification output. A content node does not identify which of its information is verification-significant. The significance of its contents is established by instruction nodes or by the relationship between multiple content nodes.

Node Types

Hybrid Node

A context node that functions as both instruction and content. A document that says "Your standard is WCAG 2.2 AA" (content) and also says "Always verify your standard against the current published version" (instruction to check another node) is a hybrid node. Hybrid nodes create denser web topologies.

Node Types

Document Stack

An ordered set of context nodes loaded sequentially by an agent during its boot or execution sequence. The document stack is the web in its temporal form.

Security Properties

Node Leak Resilience

The property that leaking any single context node does not reveal the verification output, because the output depends on relationships between nodes that the leaked node does not describe.

Security Properties

Fabrication Detection

The failure mode detection capability in which an AI agent's plausible but incorrect verification output is caught because the multi-node structure makes the fabrication detectable. The richer the web topology, the harder fabrication becomes.

Security Properties

Rotation

The act of changing any node in the web to invalidate the current verification output. In an N-node PassWeb, there are N axes of rotation. Each axis is independent.

Security Properties

Comprehension-Alignment Coupling

The property that context nodes contain both verification-significant information and governing directives. Running verification loads those directives into context alongside verification-significant content. A pass indicates the expected verification output matched; it does not guarantee downstream obedience to every behavioral rule.

Security Properties

Embedded Instruction

Information within a context node that functions as an instruction to the agent, but only if the agent has already loaded other specific nodes. Without those other nodes, the embedded instruction reads as ordinary content.

Security Properties

Version Attestation

Because the verification output depends on all nodes, a stale version of any single node produces an incorrect output. A single verification check confirms that all N documents are current.

Failure Modes

FM-1: False Negative (Karl-Type)

Agent lacks content nodes. Fabricates output from training data. Detection: fabricated output does not match expected.

Failure Modes

FM-2: False Positive (Assembly-Type)

Agent passes PassWeb at one level, fabricates at a lower level where no PassWeb exists. Mitigation: Layered PassWebs at every level.

Failure Modes

FM-3: Full Web Leak

Adversary obtains all nodes. Can produce verification output. Mitigation: Separate storage, independent access controls, frequent rotation.

Failure Modes

FM-4: Partial Web Leak

Adversary obtains some but not all nodes. Cannot produce verification output if the missing nodes matter to the configured relationship. Resilience depends on which nodes leak, not only how many.

Failure Modes

FM-5: Training Data Collision

Verification information exists in training data. Agent might guess correctly. Mitigation: Use novel, invented information.

Failure Modes

FM-6: Context Window Overflow

Web requires more tokens than agent can hold. Agent cannot process all nodes. Mitigation: Design within context window limits.

Failure Modes

FM-7: Context Injection Attack

Adversary injects fake instruction node mid-session. Mitigation: Instruction nodes via injection-resistant channels.

Failure Modes

FM-8: Partial Document Load

Content node loaded incompletely. Verification output incorrect. Detection: Fails verification.

Failure Modes

FM-9: Output Format Variation

Agent produces correct output in unexpected format. Mitigation: Normalize output before comparison.

Failure Modes

FM-10: Web Topology Inference

Adversary infers web structure from agent behavior. Severity: Low. Knowing topology without node contents does not enable producing verification output.

Failure Modes

FM-11: Attention / Skimming Failure

All required nodes fit in the context window, but the model under-uses middle or low-salience content (lost in the middle). Verification can still fail when the expected output requires combining those pieces.

Chat with AI PassWeb

Conversation

Getting started

Ask me anything about AI PassWeb, context verification, split-context architecture, or how the mechanism works.