What's happening right now? Search memories, inspect active gates, manage your team, and export training data. This is your control plane for AI agent behavior.
🔍 Search — find any memory🛡️ Gates — what's blocking👥 Team — org metrics📤 Export — DPO training data
📊 Demo Mode — sample data. Pro unlocks your personal dashboard with search, DPO export, and gate analytics.Start 7-day free trial
No review checkpoint yet. Mark the current dashboard as reviewed to start seeing only new changes.
New feedback
0
events since checkpoint
New negatives
0
fresh review risk
New lessons
0
promoted since checkpoint
New blocks
0
gate hits since checkpoint
Checkpoint
Waiting for dashboard data...
Newest items
No new review activity yet.
🔍 Search Memories
🛡️ Active Gates
👥 Team
🧩 Generated Views
⚙️ Policy Origins
🧱 Gate Templates
📊 Insights
📦 Export
Enter a query to search your memories
Active Pre-Action Gates
Loading gates...
Shared Team Reliability
See which agents drift, which gates save the team the most time, and whether the shared workflow is getting safer.
Loading team metrics...
Loading predictive insights...
Highest-Risk Agents
Loading agent risk...
Top Blocked Gates
Loading blocked gates...
Predictive Watchlist
Loading predictive watchlist...
Generated Hosted Views
A constrained JSON spec builds opinionated review dashboards from approved cards, lists, and callouts.
Loading generated views...
Loading generated view...
Policy Origins
Loading settings hierarchy...
Loading settings summary...
Layer Precedence
Loading layers...
Routing Preview
Loading routing...
Resolved Setting Origins
Loading settings origins...
Curated Gate Templates
Loading template library...
Loading templates...
📊 Feedback Insights & Lesson Pipeline
How your thumbs-up/down signals turn into lessons that prevent repeat mistakes.
💸 Estimated tokens saved
Computed from your real blocked-action count × 2,000 input + 600 output tokens per avoided retry, priced at a Sonnet-heavy blend (80% Sonnet 4.5 / 15% Opus 4.6 / 5% Haiku 4.5). Conservative estimate — actual savings may be higher.
—
— tokens · from — blocked calls
Feedback → Lesson Pipeline
Loading pipeline...
Feedback Trend (30 days)
Lessons Generated (30 days)
Gate Effectiveness (14-day audit)
How ThumbGate Learns
👎👍
1. You React
Thumbs-down on a bad action, thumbs-up on a good one
🧠
2. Lesson Distilled
ThumbGate extracts what went wrong and how to avoid it
🛡️
3. Gate Promoted
Repeated lessons auto-promote into pre-action gates
🚫
4. Mistake Blocked
Gates intercept the same mistake before it happens again
Export Training Data (DPO)
What is DPO?
Direct Preference Optimization is a technique for fine-tuning LLMs using human preference data. Your 👍/👎 feedback is converted into training pairs:
👍 "chosen" — the response that worked
👎 "rejected" — the response that failed
Use these pairs to fine-tune any model (OpenAI, Llama, Mistral) so it actually learns from your corrections — not just blocks mistakes, but stops making them.