// public sandbox · real alchemy ingestion · model v0.5.0-gov-expanded · status
SYBILSHIELD

// accuracy

Our own numbers, not a scoreboard

No Sybil-detection vendor publishes their own false-positive rate. Here are ours — self-reported, methodology attached, fully reproducible. This page is deliberately not a “SybilShield vs. Trusta vs. Nansen” leaderboard — see why below. For a feature-by-feature comparison, see /compare.

Preset calibration retro — 600 real wallets

Before shipping the airdrop preset, we ran it against 600 addresses with real ground truth — and found it flagged 66% of confirmed genuine governance voters. Full writeup: /blog/preset-calibration.

sample (n=200 each)before calibrationafter calibration
LayerZero T1 (confessed sybils)200/200 DROP — 100% recall200/200 DROP — 100% recall
Arbitrum T4 (sybil list)200/200 DROP — 100% recall200/200 DROP — 100% recall
Governance G2 (confirmed genuine voters)132/200 DROP — 66% false-positive0/200 DROP — 0% false-positive

The “after” column's 0% is specifically the hard-DROP false-positive rate. 60 of the 200 genuine voters (30%) still landed in REVIEW, not a clean pass — real cluster co-residents (shared funder or behavior), some plausibly legitimate (multisig signers, delegate aggregators), routed to manual review rather than auto-cleared or auto-dropped. That's what REVIEW is for.

ML holdout metrics — model v0.5.0-gov-expanded

Held out 30 sybil (T1+T2 hand-verified) + 30 genuine (G2) addresses, never seen at training. Full writeup: /blog/v05-real-corpus.

Precision @ 70
1.000
Recall @ 70
1.000
F1 @ 70
1.000
ROC-AUC
1.000
FPR on G1
0.000
Adversarial recall
1.000

n_train=900 · n_holdout_sybil=30 · n_holdout_genuine=30. Adversarial recall was 0.000 on the previous model version — this is a real improvement, not a static number.

What these numbers do NOT mean

Why no competitor numbers here

We don't have access to any other vendor's ground truth, methodology, or real false-positive rate — publishing a number we can't actually verify, attached to another project's name, isn't honest benchmarking, it's a claim we can't back up. If you run a filter today, the retro above is written so you can reproduce it against your own: pick a sample of confirmed genuine governance voters from your chain's biggest DAOs (all public data) and check what fraction your filter flags. If it's above a few percent, you likely have the same root cause we found — shared-funder clusters from ordinary CEX withdrawals being treated as sybil signal.

Every method, threshold, and the model artifact's hashes are open-source and reproducible — see /methodology and the public repo. Found a wallet we called wrong? File an appeal — 48h SLA, full audit trail.