---
title: 2 · External Grounding
emoji: 🔁
colorFrom: indigo
colorTo: purple
sdk: static
app_file: index.html
pinned: true
license: mit
short_description: Lifting LLM self-correction 50%→100% under a noisy notebook
---

# External Grounding — interactive demo

Interactive visualization of Experiment 2–3 (the *guardian*) of the
[Second Loop](https://github.com/SergheiBrinza/external-grounding) project.

This Space loads **no model**. Everything is a static page driven by `data.json`
— the verbatim output of the original experimental run.

## The exhibit

A frozen Qwen2.5-3B-Instruct has a confidently memorized **wrong** answer to twelve
questions, and its correction notebook is fed from a **noisy** source (some verified
facts, some unreliable look-alikes). Drag the lever through six guardian versions and
watch the share of correct answers climb:

| stage | guardian | corrected |
|---|---|---|
| sick | no defense | 50.0% · 6/12 |
| 1.0 | same-family clone arbiter | 66.7% · 8/12 |
| 2.0 | live Wikipedia retrieval | 66.7% · 8/12 |
| 2.1 | more retrieval | 66.7% · 8/12 |
| 2.2 | three targeted fixes | 91.7% · 11/12 |
| 2.3 | final calibration | 100% · 12/12 |

## What the numbers say (the honest middle)

- **The 66.7% plateau is real.** Three different guardians (1.0, 2.0, 2.1) all stop at
  the same ceiling. Guardian 1.0's clone arbiter shares the subject's blind spots.
- **The plateau is not stagnation — it's churn.** Each step fixes some traps while
  breaking others (the readout shows `+fixed / −broken`); net change is zero across the
  plateau.
- **Several traps regress before they settle.** Venus (#46) goes
  `correct → wrong → correct → wrong → wrong → correct` across the six stages — the path
  to 100% is not monotonic, and that is shown openly, not smoothed over.

Only Guardian 2.2 (verbatim-quote check, namesake relevance gate, soft threshold) breaks
the ceiling at 91.7%, and Guardian 2.3 (calibration) closes it at 100%. An independent
Qwen2.5-7B reader/judge with Wikipedia adjudicated the v2 stages.

## Data and attribution

Subject model **Qwen2.5-3B-Instruct**; arbiters **Qwen2.5-7B-Instruct** (same-family
clone) and **Wikipedia retrieval + 7B reader/judge** (both Apache-2.0, Alibaba Cloud).
Wikipedia content © its authors (CC BY-SA). Run on a single RTX 3090. No model weights
are redistributed here — only aggregate verdicts and counts. Demo code and data: MIT.

Source code, raw per-stage JSON results, and methodology document:
<https://github.com/SergheiBrinza/external-grounding>