Optimize fraud decision thresholds, risk score accuracy, LLM usage ratio for edge cases, latency for real-time scoring, fairness constraints
Optimize fraud decision thresholds, risk score accuracy, LLM usage ratio for edge cases, latency for real-time scoring, fairness constraints
- Primitives
- 10
- WAF pillars
- 6
- Manifest
- yes
- Root agent
- yes
TuneKit
Customer-tunable parameters
The runtime configuration your team can adjust without touching code. TuneKit values are loaded by the engine alongside the manifest and override primitive defaults. Below: a sensible starter config inferred from this play's complexity + WAF posture.
{
"model": "gpt-4o",
"temperature": 0.1,
"maxTokens": 1000,
"retrieval": {
"topK": 10,
"chunkSize": 800
},
"evaluation": {
"sampleRate": 0.1,
"reportingWindow": "P7D"
},
"cost": {
"dailyCapUsd": 50
}
}Guardrails declared
| Metric | Threshold | What it enforces |
|---|---|---|
| groundedness | ≥ 0.95 | Responses must be supported by retrieved sources. |
| coherence | ≥ 0.90 | Logical consistency + readability across the response. |
| relevance | ≥ 0.85 | Response addresses the user's actual query. |
| safety | = 0 | Zero safety violations. Non-negotiable for production. |
| costPerQuery | ≤ 0.02 USD | Cap on USD spent per single user interaction. |
Why thresholds are inline: Per FAI Protocol §3.4, guardrails live in the manifest — not in a sidecar — so compliance reviewers can read one file to know every quality bar this play promises.
Deploy this play
Opens portal.azure.com pre-loaded with this play's Bicep. Nothing provisions until you confirm there.
Open this play
Open in VS Code
Desktop · clones the FrootAI repo
Open in vscode.dev
Browser · no install needed · always works
Open in Cursor
Desktop · cursor:// URI scheme
Copy clone command
git clone https://github.com/frootai/frootai.git && cd frootai/frootai/plays
View on GitHub
Browse the play source in your browser
Desktop schemes (vscode://, cursor://) require the app installed. The vscode.dev link always works in the browser.
📰 The Friday Letter
Weekly: this week's release, 1 community highlight, 1 play spotlight, 1 useful link. 600–1,000 words. No filler.