What's happening right now? Search memories, inspect active checks, 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 Pro now
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 Checks
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...
Agent Surface Inventory
See which tools, policy sources, and MCP surfaces are actually active before you widen rollout.
Loading inventory...
Observed Tools
Loading observed tools...
Policy Sources
Loading policy sources...
Background Agent Mode
Queued and scheduled agents need explicit checkpoints before they touch code, money, or customer systems.
Loading background-agent mode...
Governance Pressure
Loading background-agent governance...
Run Types
Loading background-agent run types...
Regulated Buyer Proof
Show where policy came from, how many runs were reviewed, and which proof artifacts are ready when buyers ask for auditability.
Loading proof posture...
Policy Origin Proof
Loading policy origin proof...
Latest Replay Artifacts
Loading replay artifacts...
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)
Highest-ROI Next Actions
These recommendations come from real feedback, risk patterns, and delegation pressure in your current runtime.
Loading next actions...
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 checks
🚫
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.