Agent Trust Oracle — Methodology (v2, cross-chain)

Read-only trust scores for ERC-8004 agents across multiple EVM chains. Default chain is ethereum. Wired chains: base, bsc, ethereum, mantle, polygon. Values below are pulled live from scoring.py.

Score range

0 to 100. Or null when data is too thin (see refusal threshold).

Components & weights

Refusal threshold (insufficient_data)

If the agent has fewer than 3 distinct clients, the response is score = null with status = "insufficient_data". Single largest sybil-mitigation lever.

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

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

Log-axis reference points

• 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

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

Half-life: 50,000 blocks. Block-rate varies per chain; we still apply the same half-life — older entries weigh less proportionally to that chain’s block velocity.

Final aggregate

score = value_avg * 0.50
      + client_breadth * 0.20
      + volume * 0.15
      + recency * 0.15

Identity is per-chain

We score by (chain, agent_id). We do NOT merge agent IDs across chains. A probe on 2026-06-16 found zero owner wallets with agents on more than one chain, so a cross-chain operator view has no signal today. Future work-item once that changes.

What we do NOT do (yet)

• No client-graph sybil detection.
• No tag-based weighting.
• No off-chain reputation imports.
• No cross-chain agent identity merge.

Sources (read-only)

• Canonical Identity/Reputation registries at 0x8004A169FB4a3325136EB29fA0ceB6D2e539a432 + 0x8004BAa17C55a88189AE136b182e5fdA19dE9b63 on every wired chain.
• Source code: github.com/Nikoble1926/agent-trust-oracle
• Tamper-evidence: /provable/verify + /provable/head. Each entry stores per-chain counters + a agents_sha256 digest of the full sorted per-agent table (not the table itself). Anyone can recompute the table from data/<chain>/feedback.jsonl and check the digest with python3 provable.py recompute <chain>. Chain mirrored to github.com/Nikoble1926/agent-trust-oracle/blob/main/provable/scores_chain.jsonl.