Skip to content
GLACIS GLACIS Documentation

Epochs & Sampling

GLACIS organizes attestations into discrete time periods called epochs and uses sampling to balance coverage with storage.

Epochs

An epoch is a bounded time period (default: 1 hour):

interface Epoch {
  id: string;           // e.g., 'epoch_2024010112'
  startTime: number;    // Unix timestamp (ms)
  endTime: number;      // Unix timestamp (ms)
  duration: number;     // 3600000 (1 hour)
  attestationCount: number;
  merkleRoot?: string;  // Set when epoch closes
}

Why Epochs?

  • Ordering: Attestations are ordered within epochs
  • Freshness: Bearer tokens expire with epochs
  • Batching: Auditors can verify entire epochs
  • Control: Sampling rates can vary per epoch

Sampling

L0 vs L2

LevelGeneratedContentsSize
L0Every requestMetadata only~200 bytes
L2SampledFull evidence~2-10 KB

Sampling Rate

sampling: {
  rate: 100,  // 1 in 100 requests get L2
}

Adaptive Sampling

rules: [
  // Always sample GPT-4
  { condition: { field: 'model', operator: 'eq', value: 'gpt-4' }, rate: 1 },
  // Higher rate for long prompts
  { condition: { field: 'content_length', operator: 'gt', value: 2000 }, rate: 10 },
]

Epoch Lifecycle

  1. Epoch starts (e.g., 12:00:00)
  2. Bearer tokens issued for this epoch
  3. Attestations collected
  4. Epoch ends (e.g., 13:00:00)
  5. Merkle tree finalized
  6. New epoch starts

Next Steps