Auditors cannot read every contract. ~95% of EVM mainnet has no verified source. Static analyzers like Slither and Aderyn — both brilliant tools — cannot see what they cannot compile. We built REVERT to score what they cannot reach: bytecode-only ML triage at ~1 ms per contract on a single CPU core (~0.97 ms feature extraction + ~0.04 ms inference; ~106 contracts/hour as a throughput envelope, not a guarantee; sub-second end-to-end including the on-chain bytecode fetch).
A high score = review priority: the score tracks attack-surface and complexity, so complex/audited infra like DEX routers, lending pools and bridges also scores high. A high score is NOT a confirmed vulnerability, and a low score does NOT mean safe. REVERT is the bytecode-only first-pass complexity / attack-surface filter for the contracts Slither can't open — calibrated to agree with Slither's contract-level signal on verified contracts — triage, not audit — and metrics are upper-bounded by Slither's own label quality and not yet validated against confirmed exploits on the unverified mainnet domain it scores.
Smart-contract security is upstream of every DeFi outcome. Yet most security tools assume source code is available. Most of mainnet does not. The gap is real and measurable, and bytecode-derived complexity is a fair target for ML triage.
Our gradient-boosted classifier (70 hand-engineered features, trained on
117,091 Slither-labelled contracts) — the served model,
xgb-2025-12-28 — reaches F1 ≈ 0.92–0.95 for agreeing with
Slither's binary contract-level flag on 11,670 held-out, source-verified
contracts (95% bootstrap CI [0.918, 0.948]; tree ensembles RandomForest /
XGBoost are statistically indistinguishable). This is agreement with
Slither on verified contracts, not measured accuracy on the unverified
mainnet contracts REVERT actually triages — the two differ by distribution
shift.
On the multi-label SWC task (8 classes) the same model reaches macro-F1 = 0.75 for reproducing Slither's per-class labels, with roughly 70-90% per-class precision; rare classes (bad-randomness, double-spending) are the hardest, bounded by Slither label scarcity. These are agreement- with-Slither scores on verified contracts, not exploit-confirmed accuracy. Full methodology, ablation, and code are open — see the research paper & dataset (presented at ICICPE 2026, a computer-science / education venue).
Sergei Solovev — researcher at HSE Faculty of Computer Science, Moscow. Background in applied ML for finance (LOB mid-price prediction) and smart-contract vulnerability detection. Published preprints on figshare; lead author of the ICICPE 2026 paper that underpins REVERT. WorldQuant University scholarship.
Contact: ssolovev@revert.pro
REVERT is the ML triage layer that prioritises which contracts feed a three-tier review stack:
We compose — not replace. Our value is funnel narrowing: from 1 000 contracts to the 10 that deserve a human auditor's time.
Atlas Trading Agency LLC (Trade Agency Atlas
LLC), operating the REVERT brand. Registered in Bishkek,
Kyrgyzstan on 2025-11-03. Registration №321952-3301-OOO,
INN 00311202510097, 55 Logvinenko St., Office 243, Bishkek, Kyrgyzstan. Sole owner
and director: Sergei Solovev. Public registry:
record.minjust.gov.kg/view/00311202510097.
Bootstrapped, no outside funding.
forge revert <address>)See /changelog for release notes.
Our edge compounds over time — the proprietary bytecode dataset, deepening tool integrations, and direct auditor relationships get stronger every month. Bootstrapped means no investor pressure to upsell features you don't need; high-margin, customer-funded, built for the long run. We may raise later only if a specific customer wedge needs it.