External Grounding
Raising self-correction from 50% to 92.9% under a noisy notebook
Static demo · no live model
★  Second Loop · Part 2 of 3  ★
Subject modelQwen2.5-3B-Instruct (frozen)
Arbiter v1Qwen2.5-7B (same-family clone)
Arbiter v2Wikipedia + 7B reader
LicenseMIT
No defense — raw 3B
50.0%
7 / 14 on the noisy notebook
Clone-arbiter ceiling
64.3%
= the arbiter's own accuracy
Calibrated — Guardian 2.3
100%
14 / 14, external grounding
incl. one trap (#46) disambiguated; without it 92.9% (13/14)
Drag the guardian from sick to calibrated.
Move the lever through six guardian versions and watch corrected answers climb 50% → 92.9%, and to 100% once one ambiguous trap (#46) is reformulated
Current guardian stage
Sick (no defense)
stage 1 of 6
Corrected answers · higher is better
50.0%
7 / 14 traps correct
No external grounding
vs previous stagestart
net change
The lever — guardian version ← weaker  ·  stronger →
No defense 64.3% clone ceiling Calibrated
All 14 traps — colour = verdict at the current stage · dots = full trajectory · click to inspect Stage: 7 / 14 correct
#46 science-number
Correct answer
Memorized wrong answer
Final answer · Guardian 2.3
Wikipedia retrieval — top sources
trust v2.2 trust v2.3
Verdict across the six guardian stages