Agent Trust Oracle — Methodology (v1)

Read-only trust scores for ERC-8004 agents on Base mainnet. This page is the canonical reference for the scoring math; the values below are pulled live from scoring.py so they can never drift from the implementation.

Score range

0 to 100. Higher = more positive feedback from more independent clients, recently. Or null when data is too thin (see refusal threshold below).

Components & weights

Each component is computed independently on a 0–100 scale and then combined with the fixed weights below. Sum of weights = 1.00.

Refusal threshold (insufficient_data)

If the agent has fewer than 3 distinct clients leaving feedback, we return score = null with status = "insufficient_data — fewer than 3 distinct clients have left feedback". This prevents one client farming a high score by self-rating and is the single largest sybil-mitigation lever in v1.

Value normalisation (−100..100 → 0..100)

ERC-8004 feedback values are signed decimals with arbitrary scale (e.g. 99.77, 5.0, -3.0). We clamp to the inclusive range [-100, +100] then linearly remap to [0, 100]:

normalised_value = (clamp(value, -100, +100) + 100) / 2

Log-axis reference points

Distinct-client breadth and total feedback volume are mapped via log1p to the 0–100 range. Saturation references:

• CLIENT_BREADTH_REF = 25.0 distinct clients → component saturates at 100.
• VOLUME_REF = 50.0 feedback entries → component saturates at 100.

def log_axis(count, ref):
    if count <= 0: return 0.0
    return min(100, 100 * log1p(count) / log1p(ref))

Recency decay

Each feedback entry is weighted by an exponential decay on chain-block age:

recency_weight(block) = 0.5 ** ((latest_block - block) / 50000)

Half-life is 50,000 blocks (~27.8 hours at Base’s 2-second block time). Half of a year-old feedback is roughly worth 0.5 ** (315.4) ≈ vanishingly small.

Final aggregate

score = sum(component[k] * WEIGHTS[k] for k in WEIGHTS)
      = value_avg * 0.50
      + client_breadth * 0.20
      + volume * 0.15
      + recency * 0.15

What we do NOT do (yet)

• No client-clustering / address-graph sybil detection.
• No tag-based weighting (the indexedTag1 & tag1/tag2 strings are passed through in raw feedback but do not bias the score).
• No off-chain reputation imports (Twitter, GitHub, etc.).
• No agent-supplied URI scraping — agent_uri on Registered events is surfaced verbatim, never trusted.

v2 work-items would tighten the sybil model and add tag-conditional scoring. For now we prefer to refuse cleanly when data is thin (see threshold above).

Sources (read-only, Base mainnet)

• IdentityRegistry   0x8004A169FB4a3325136EB29fA0ceB6D2e539a432
• ReputationRegistry 0x8004BAa17C55a88189AE136b182e5fdA19dE9b63
• Source code: github.com/Nikoble1926/agent-trust-oracle
• Tamper-evidence: /provable/verify (re-hashes the snapshot chain) and /provable/head (latest entry). The chain itself is mirrored to github.com/Nikoble1926/agent-trust-oracle/blob/main/provable/scores_chain.jsonl so every push is a third-party timestamp anchor.