Credit Integrity
How DAISI keeps the credit system fair without slowing down inference.
Overview
DAISI credits are earned by hosts that process inference requests and spent by consumers who use the network. Because hosts self-report token counts and the orchestrator (ORC) awards credits in real time, the system is designed around trust — but trust needs verification. DAISI uses an async anomaly detection system that runs entirely in the background, so inference speed is never affected.
The core principle: never add latency to the inference hot path. All anti-gaming measures are after-the-fact detection and human review, not inline blocking.
Receipt Deduplication
Every inference receipt is tracked by the ORC using a fast in-memory lookup keyed by host ID and inference ID. If the same receipt arrives twice, the duplicate is silently dropped and logged. This prevents a host from submitting the same completed inference multiple times to earn duplicate credits.
-
Zero latency impact A single dictionary lookup — the fastest possible check.
-
24-hour window Entries expire automatically after 24 hours to keep memory usage bounded.
Anomaly Detection
A background service scans for suspicious credit patterns every 30 minutes. When a pattern is detected, it creates an anomaly record for admin review. No automatic penalties are applied — a human always makes the call.
What gets flagged
Inflated Token Counts
If a host reports an average token count per inference that's more than 10x the network-wide median, it's flagged. This catches hosts that inflate numbers to earn more credits per request.
Receipt Volume Spikes
If a host suddenly submits 3x more receipts in an hour than its 7-day hourly average, it's flagged. This catches hosts that artificially generate traffic to earn credits.
Zero-Work Uptime
Hosts that stay online for 7+ days without processing any inferences are surfaced to admins. This isn't necessarily malicious — some hosts may be in low-demand regions — but it's worth reviewing.
Circular Credit Flow
If two accounts consistently serve each other's inference requests, they may be colluding to generate artificial credit flow. Both accounts are flagged when this pattern is detected.
Admin Review
All anomalies are reviewed by DAISI administrators through the Manager admin panel. Admins can dismiss false positives or take action on confirmed gaming. If credits need to be adjusted, they use the existing credit adjustment tools to correct balances.
This human-in-the-loop approach ensures that legitimate hosts are never automatically penalized while still catching bad actors quickly.
Tool Execution Billing Integrity
Secure marketplace tools that charge per-execution credit costs have their own set of integrity protections. These ensure that consumers are charged fairly and that providers cannot manipulate billing.
Protections
Session Validation
Every tool execution requires the provider to validate the consumer's session with the ORC before executing. The ORC verifies that the session is active and that the tool is entitled. This prevents unauthorized execution and ensures that only legitimate sessions incur charges.
Cost Snapshot
The per-execution credit cost is snapshotted at the time of execution and recorded immutably with the execution record. If a provider changes their tool's price after an execution, the original cost stands. This prevents retroactive price manipulation and ensures consumers are always charged the price that was in effect when they used the tool.
Background Aggregation
Execution charges are not deducted in real time (which would add latency to tool calls). Instead, a background billing service aggregates executions every 60 seconds and processes them as batched credit deductions. This keeps inference fast while ensuring all charges are eventually applied.
For Hosts
If you're running a legitimate host, this system works in your favor. By detecting and removing bad actors who game the credit system, honest hosts get a fair share of the credit pool. You don't need to do anything differently — just keep your host running and processing real inference requests.
If you receive a notification about an anomaly flag, it doesn't mean you've done anything wrong. Admins review every case individually and will reach out if clarification is needed.