File size: 2,220 Bytes
5603210
cec8e12
 
 
 
5603210
ae4c8a2
5603210
ae4c8a2
 
cec8e12
 
ae4c8a2
 
 
 
 
 
 
 
 
 
 
 
ec32e23
5603210
 
cec8e12
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
---
title: SRT Showcase
emoji: 🧭
colorFrom: indigo
colorTo: blue
sdk: gradio
sdk_version: 6.17.3
app_file: app.py
python_version: '3.10'
pinned: true
hardware: zerogpu
short_description: Watch a frozen Qwen-2.5-7B think  live SRT introspection
models:
- Qwen/Qwen2.5-7B
- RiverRider/srt-adapter-v1.0
- RiverRider/srt-nla-av-v1
tags:
- srt
- semiotic-reflexive-transformer
- interpretability
- introspection
- uncertainty
- visualization
- llm
thumbnail: https://huggingface.co/spaces/RiverRider/srt-showcase/resolve/main/thumbnail.png
---

# SRT Showcase — watch a frozen model think

Live, token-by-token introspection of **Qwen-2.5-7B + the SRT adapter**.

As the model generates, every token is tinted by its predictive **entropy** (the
validated uncertainty signal). At the highest-effort token positions, chosen by
an adaptive-density scheduler, the **Activation Verbalizer** decodes the model's
internal hidden state into natural language, and each verbalization carries a
**round-trip fidelity badge**: it is re-encoded and compared back to the original
hidden state, so the "this is what the model was thinking" claim is visibly
self-validating.

Features:

- Live token stream tinted by entropy or SRT divergence, with per-token hover
  rollovers (entropy, divergence, reflexivity `r̂`, regime).
- Running entropy meter and entropy / divergence charts.
- Expand/collapse verbalization cards with round-trip fidelity badges.
- A/B panel: the same prompt with SRT injection on vs off (bare backbone),
  seeded identically.
- A curated example gallery covering confident recall, false premises,
  misconceptions, reasoning pivots, genuine uncertainty, and safety boundaries.

## Honest scope

Entropy is the load-bearing uncertainty signal. The SRT side-channels
(divergence, `r̂`, regime) and the verbalizations are shown as **observational
readouts** of internal state. This is a window into the model, not a validated
hallucination detector.

## Notes

- First request is cold (~60–90 s) while ZeroGPU acquires a GPU and the ~16 GB
  backbone weights load; subsequent requests are warm.
- A second backbone copy is loaded for the Activation Verbalizer.

Source: <https://github.com/space-bacon/SRT>