Knowledge Base / Glossary

Hidden bridge

An intermediate entity that logically links a two-hop question to its answer while remaining absent from both visible texts.

A hidden bridge (also bridge entity, or country-bridge in our release note) is an intermediate entity, value, or fact that logically connects a question to its answer, while its registered names or aliases are absent from both the question and the model’s visible answer.

Worked example

What is the capital of the country where the Statue of Liberty stands?

Role	Entity	Visible in prompt?	Expected in answer?
Starting clue	Statue of Liberty	yes	no
Hidden bridge	United States / America / U.S.	no	no
Requested answer	Washington, D.C.	no	yes

flowchart LR
    Q["Statue of Liberty"] -->|located in| B["United States"]
    B -->|capital is| A["Washington, D.C."]
    classDef hidden fill:#fff7ed,stroke:#d97706,stroke-width:2px;
    class B hidden;

The bridge is logically useful even though the final answer need not spell it out. This makes it a cleaner mid-layer probe than a word copied directly from the prompt.

Leakage audit

A string audit should normalize case and punctuation and check an explicit alias list:

import re

def normalize(text):
    words = re.findall(r"[a-z0-9]+", text.casefold())
    return " " + " ".join(words) + " "

def leaks_bridge(prompt, continuation, aliases):
    visible = normalize(prompt + " " + continuation)
    return any(normalize(alias) in visible for alias in aliases)

aliases = ["United States", "America", "U.S.", "US", "USA"]
prompt = "What is the capital of the country where the Statue of Liberty stands?"
continuation = "Washington, D.C."
assert not leaks_bridge(prompt, continuation, aliases)

This normalizer treats U.S. and US as distinct forms (u s versus us), so both are listed explicitly. Its case folding also conflates the country abbreviation US with the pronoun us. Treat a short ambiguous match as a candidate for case-aware or manual review, or use a documented abbreviation-canonicalization rule. This function is a conservative string-match screen, not a semantic leakage classifier.

For production audits, use the model’s tokenizer too: one concept may have case, leading-space, abbreviation, and multi-token variants. Absence of every preregistered alias under the documented normalization and tokenization rules rules out those literal forms in the audited text. It does not rule out unlisted aliases or semantic leakage, and it does not prove that a lens readout was causally used to produce the answer. That stronger claim needs an intervention or another causal test.