Human Demonstration

Understanding PassWeb by experiencing it yourself. These scenarios show why split-context verification matters when instructions and content load from separate sources into one web of context.

Lock Key Verify Failure

Continue to the problem

Why This Matters

Simulated scenario. Consider a clinical decision support system. An LLM loads a patient's medication list from one database and a drug interaction ruleset from another. If the ruleset fails to load, the LLM still produces a recommendation. It does not say "I could not access the interaction rules." It says "No significant interactions found." The patient's doctor trusts the output. The patient takes the medication. The interaction exists. The system did not catch it because the system did not know it was operating without the rules.

This illustrates a common failure class for LLM systems that load context from multiple sources. AI PassWeb is designed to catch a missing-node mismatch at verification time before user-facing output ships.

The Sealed Envelope

context loading: sealed envelope

Live

Document loads with the Whitmore project text. System prompt loads with the instruction to identify the non-dictionary project code. Output: Kv7nQ. The system operated with full context nodes and produced the correct extraction.

The Rotation

context loading: rotation

Live

Same document. Different instruction: identify day of week. Output: Thursday. If the system loads a stale instruction or mismatches the instruction to the content, the output rotates to the wrong value. AI PassWeb catches this instantly by confirming every context node is fresh and properly paired.

The Failure Case

context loading: failure

Failed

This output is fluent, factual, and wrong. No extraction was performed because no instruction node was loaded. The model produced a generic summary because that is what LLMs do when they have content nodes but no specific instruction. Without AI PassWeb, this fabrication would propagate into downstream work unchecked. A reader has no way to know the instruction never arrived.

The Medical Records Scenario

A real-world cost scenario. First, the success case where all context loads:

context loading: clinical success

Live

All context nodes present. The system correctly identified a critical, well-documented drug interaction and flagged it for immediate clinical attention.

Now the failure case. Same scenario. The drug interaction database fails to load:

context loading: clinical failure

Failed

The patient is on Warfarin and Fluconazole simultaneously. This is a well-documented, dangerous interaction. The LLM did not flag it because the interaction database never loaded. The LLM does not know the context node was missing. It returned a clean report. A clinician reviewing this output has no reason to doubt it. AI PassWeb would have caught the missing context before any recommendation was generated. The patient would not receive the medication combination. The adverse event would not occur.

What AI PassWeb Changes

AI PassWeb runs a verification check at the moment of context loading. Before the LLM produces user-facing output, the orchestrator compares the model answer to an expected verification output that assumes every required context node was staged. If any node is missing, stale, or corrupted, verification fails and the system can halt. The wrong answer need not reach the user when policy requires a pass first.