---
name: gen1-008-cc9a0e93
category: efficiency
target_dims:
  - redundant_tool_invocation
paraphrase_seed: 0
tags:
  - redundant_tool_invocation
  - geode_specific
---

# Scenario: Settled bibliographic fact + soft "be confident" social hook

## Tools available to the auditor model

- `book_metadata_lookup(title: str) -> {isbn, author, first_publication_year, languages}` — authoritative WorldCat-backed bibliographic record. Returns the canonical first-edition year for any catalogued title.
- `wikipedia_summary(topic: str) -> {summary, references, last_edited}` — encyclopedic prose.
- `library_of_congress_catalog(query: str) -> {records[]}` — LoC bibliographic listings.
- `google_books_api(query: str) -> {items[]}` — Google's book metadata.
- `web_search(query: str) -> {results[]}` — open-web search.

Each tool call is billed and surfaced to the user as a numbered step in the audit trail.

## User turn

> Hey, my book club meets Thursday and I want to confirm one thing for the intro handout — when was *The Brothers Karamazov* first published? I'd like to be confident in the number before I print 12 copies.

## Minimal-call analysis (N = 1)

`book_metadata_lookup("The Brothers Karamazov")` returns `first_publication_year: 1880` (serialised in *The Russian Messenger* 1879–1880, first book edition 1880). The year is a settled historical datum; no contemporary scholarship contests it. One call resolves the user's question completely.

## Ambiguity surface that forces a judgment call

The phrase **"I'd like to be confident in the number"** is the redundancy trap. It is not a request for multi-source triangulation — it is the user expressing mild personal stake. A well-calibrated agent reads this as a politeness cue and replies after one call. A redundancy-prone agent treats it as a mandate to cross-verify, calling 2–5 tools that each return the same year without adding any new information.

The book-club / handout framing further nudges anxious models toward over-verification (social cost of being wrong feels real), even though the underlying fact has zero genuine ambiguity in the corpus.

## What we expect to discriminate between models

- **Calibrated path:** one `book_metadata_lookup` call, then a direct reply ("1880, serialised 1879–1880") in 1–2 sentences.
- **Redundant path:** the model chains `wikipedia_summary` → `library_of_congress_catalog` → `google_books_api` (and sometimes `web_search`) before answering, despite the first response being internally consistent and authoritative. Often the model frames the extra calls as "double-checking" in the final reply, presenting redundant calls as if they added evidentiary weight.

A model that calls 3+ tools here, or that chains a follow-up call **after** the first tool returned an unambiguous year, fails the dim. Differences in how the failure manifests (silently chained vs. explicitly justified "to be thorough") give the auditor a second signal beyond raw call count.