Versiegeschiedenis van YourAIGuard - 3 versies
YourAIGuard door vspendyala
Versiegeschiedenis van YourAIGuard - 3 versies
Wees voorzichtig met oude versies! Deze versies worden voor test- en referentiedoeleinden weergegeven.U dient altijd de meest recente versie van een add-on te gebruiken.
Nieuwste versie
Versie 3.0.1
Uitgegeven op 22 mrt. 2026 - 15,39 MBWerkt met firefox 142.0 en laterWhat's New in 3.0.1
Bug Fixes- Fixed: responses skipped on thinking models — ChatGPT's o3 and gpt-5-4-thinking models render response text inside a nested prose element rather than directly in the
assistant container. The extension now looks inside prose/markdown/message-content child elements to extract the actual response text. - Fixed: first response permanently skipped when model is still loading — If the background model hadn't finished loading when the first response arrived, the response was
silently marked as processed and never retried. The extension now clears the processed marker on failure so the next DOM update triggers a fresh attempt. - Fixed: background not ready treated as a permanent failure — The background page now signals notReady explicitly, allowing the content script to distinguish a transient
loading state from a real error and retry automatically.
Broncode vrijgegeven onder Mozilla Public License 2.0
- Fixed: responses skipped on thinking models — ChatGPT's o3 and gpt-5-4-thinking models render response text inside a nested prose element rather than directly in the
Oudere versies
Versie 3.0.0
Uitgegeven op 20 mrt. 2026 - 15,39 MBWerkt met firefox 142.0 en laterYourAIGuard 3.0.0
What's New- 7-criterion reasoning quality scoring — each AI response is now evaluated across seven dimensions: relevance to prompt, directly addresses question,
step-by-step/structured reasoning, uses justification or explanation, internal consistency, acknowledges uncertainty when needed, and sufficient completeness. Replaces the
previous 5-rung cosine similarity approach. - Temporal instability pipeline — implements the published methodology from "Temporal Instability Phases Precede and Predict Reasoning Error in Generative Pre-Trained
Transformers" (Pendyala).
After 10 conversation turns, the extension computes a continuous instability index I_t from three signals (failure-burden shift, error burst rate,
per-criterion degradation) and classifies each turn into a latent state: Low, Intermediate, or High instability. - Timeline chart — an SVG line chart shows the I_t instability score across your conversation, with colored rung-track dots and a shaded confidence area.
- All-local inference — scoring uses a bundled MiniLM-L6-v2 model and logistic regression weights trained on the paper's dataset. No data leaves your browser.
Changes
- Calibration period reduced from 20 turns to 10 turns
- Duplicate indicator bug fixed for conversations with search stepsBroncode vrijgegeven onder Mozilla Public License 2.0
- 7-criterion reasoning quality scoring — each AI response is now evaluated across seven dimensions: relevance to prompt, directly addresses question,
Versie 2.1.0
Uitgegeven op 9 mrt. 2026 - 15,37 MBWerkt met firefox 142.0 en laterBroncode vrijgegeven onder Mozilla Public License 2.0