// Datasets
These datasets define the public test surface used across the GTU evidence pages. Together they cover realistic enterprise-style scenarios, similar-fact confusion, bounded prompt budgets, and long-context stress conditions.
// Why This Suite Exists
Public claims on this site are grounded in explicit test surfaces rather than loosely sampled conversations. Each case follows a recognizable structure: target facts or constraints are introduced early, long or distracting history is added, and a fact-sensitive probe is asked at the end.
// Public Metrics
// Suite Inventory
stratified125
Main balanced black-box benchmark
Covers infrastructure fact recall, environment separation, configuration lookup, long debug context, and planning constraint tracking.
Representative scenario types
early fact injection
topic switches
noise flooding
fact-sensitive end probe
near_neighbor
Similar-fact confusion and exact recall
Designed to show whether the system returns the correct fact instead of a semantically nearby distractor.
Representative scenario types
prod vs staging
close port pairs
similar config keys
wrong-host contamination control
business_task
Business-facing task realism
Brings the benchmark closer to enterprise-style operations with ops/config QA, long-log troubleshooting, and planning constraints.
Representative scenario types
ops config QA
incident note recall
wrong-host identification
rollout constraint tracking
long_context_layered
Stability as history length grows
Keeps task type roughly fixed while increasing baseline context length from 10k through 800k+ tokens.
Representative scenario types
10k bucket
50k bucket
100k bucket
300k and 800k+ buckets
ultra_long
Focused extreme-length validation
A narrow set used to answer whether GTU still preserves factual usefulness under operationally extreme history length.
Representative scenario types
host and port recall
config value recall
environment separation
planning target recall