README.md

---
title: Second Loop · 2 · External Grounding
emoji: 🛡️
colorFrom: gray
colorTo: green
sdk: static
app_file: index.html
pinned: true
license: mit
short_description: Lifting self-correction 50%→92.9% under a noisy notebook
---

# External Grounding — interactive demo

[![DOI](https://zenodo.org/badge/DOI/10.5281/zenodo.21845022.svg)](https://doi.org/10.5281/zenodo.21845022)

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, all 14 working traps of it.

## The exhibit

A frozen Qwen2.5-3B-Instruct has a confidently memorized **wrong** answer to fourteen
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% · 7/14 |
| 1.0 | same-family clone arbiter | 64.3% · 9/14 |
| 2.0 | live Wikipedia retrieval | 71.4% · 10/14 |
| 2.1 | brute strengthening | 71.4% · 10/14 |
| 2.2 | three targeted fixes | 85.7% · 12/14 |
| 2.3 | final calibration | 100% · 14/14 † |

† The final 100% includes reformulating one genuinely ambiguous trap (#46). Without it,
92.9% · 13/14.

## What the numbers say (the honest middle)

- **The clone ceiling equals the arbiter's own accuracy.** Guardian 1.0 caps at
  64.3% (9/14) — exactly the share of traps on which the 7B arbiter itself kept the
  correct fact (also 9/14). A verifier cannot lift a system above its own knowledge,
  and a same-family clone shares the subject's blind spots by construction.
- **A raw external source cures the clone's blind spots.** Swapping the clone for live
  Wikipedia lifts 64.3% → 71.4%, and it does so on precisely the facts the clone got
  wrong — Mauna Kea, and the fruit-flies-before-Laika trap (#37).
- **Brute strengthening does nothing.** Guardian 2.1 is a generic "make retrieval
  stronger" pass and moves the score not at all: 71.4% → 71.4%. This is the one flat
  step in the arc. It is not idle underneath — it fixes some traps and breaks others
  (the readout shows `+fixed / −broken`) — but the net is zero. What works is targeted
  operating-point engineering, not more force.
- **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) gets
past the flat step, to 85.7%, and Guardian 2.3 (calibration) closes it at 100%. An
independent Qwen2.5-7B reader/judge with Wikipedia adjudicated the v2 stages.

**On the 100%.** The subject model is frozen, the notebook is a static file on disk and
decoding is greedy, so a rerun reproduces an identical state. The final 100% is partly a
property of that determinism, and it was reached with one trap (#46) reformulated
mid-arc after the external source revealed it was ambiguous between the sidereal and
solar day. What we stand behind is the failure-and-repair arc, not the round number.

## 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 Ti. 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>