01 / training corpus
2.1B labelled wallet behaviours
Eighteen months of normalised on-chain actions across fourteen networks, hand-labelled by a research team of seven PhDs and audited under our ISO 27001 control boundary.
Platform · v3.2 · 14 chains
Capital, allocated with intent.
PFN Dai fuses a transformer trained on 2.1 billion labelled wallet behaviours with proprietary DeFi signal models — surfacing risk roughly forty-eight hours before the market reacts, and routing capital into yield that has been audited, not asserted.
Research note · module 01
Deep Flow — the engine beneath every signal.
A proprietary transformer, retrained nightly on two billion labelled wallet behaviours. It learns to read capital the way a macro analyst reads a balance sheet — not by following price, but by modelling who is moving it, why, and at what risk.
02 / prediction target
0.87 AUC on rug-pull at T−72h
The model flags rug-pull probability seventy-two hours before the event, on an out-of-sample test set held in cold storage and never revisited after deployment.
03 / throughput
4.2M signals per second
Peak ingest rate measured on our own validator-node infrastructure across Ethereum, Arbitrum, Base, Optimism, Solana, Aptos, Sui and seven additional networks.
04 / deployment
Sub-180ms end-to-end latency
Dedicated validator nodes, co-located with sequencer RPCs, return a scored signal in fewer than 180 milliseconds — the budget the Treasury Console rebalances against.
Schedule A — technical reference
Platform specification, on a single page.
The table a procurement reviewer screenshots into a due-diligence memo. Every figure is taken from a control document, audit report, or operating log — not a sales deck.
| Domain | Item | Specification | Source / control |
|---|---|---|---|
| Coverage | Networks indexed | Ethereum, Arbitrum, Base, Optimism, Solana, Aptos, Sui, plus seven additional networks | Operating log · Jan 2026 |
| Throughput | Signal ingestion | 4.2M on-chain signals per second at peak load | Internal capacity test · Q4 2024 |
| Latency | End-to-end signal | Sub-180ms across 14 EVM and non-EVM chains | Validator-node SLO · live |
| Model | Rug-pull prediction | 0.87 AUC at T−72h on out-of-sample test set | Model card · rev. 2026.01 |
| Security | Certifications | SOC 2 Type II (Feb 2023) · ISO 27001 (Jun 2024) | Audit reports on file |
| Data residency | EU institutional clients | MiCA-aligned EU residency option, the only on-chain analytics provider with ISO 27001 plus this residency posture | Trust · policy |
| Integrations | Treasury & OMS | REST & WebSocket Signal API; read-only Treasury Console with role-based access for nine of the top fifty crypto-native funds by AUM | Client implementation log |
| Performance | APY uplift | +6.8% average APY outperformance vs. passive DeFi benchmarks over 18 months (audited; past performance is not indicative of future results) | Armanino LLP · independent audit |
Chapter II · three surfaces
Three surfaces, one engine.
The same Deep Flow transformer powers three distinct institutional products. Pick the surface that matches your operating posture — the signal underneath is identical, and so is the audit trail.
Surface 01 / analytics
Deep Flow Analytics — the research desk.
A research workbench for portfolio managers and risk officers. Pool-level rug-pull probability, capital-flow topology, and counter-party exposure are surfaced as a single scored signal, not a wall of charts. Filters by chain, asset, and venue; exports to PDF, CSV, or directly into an OMS via the Signal API.
- Pool & venue rug-pull score, refreshed every fifteen minutes
- Capital-flow topology graph with named wallet clusters
- Counter-party exposure table with concentration flags
pool · 0x71f9…3a
0.87
rug @72h · review
treasury · A21
+6.8%
APY vs. benchmark
capital flow · last 24h
# subscribe to scored risk on a venue ws.connect("wss://api.pfndai.com/v3") .subscribe("venue:0x71f9..3a") > 200 OK · latency 178ms { "venue": "0x71f9..3a", "rug_p_72h": 0.87, "tvl_usd": 412,907,221, "model": "deep_flow_v3.2" }
Surface 02 / developer
Signal API — the trading desk.
A low-latency REST and WebSocket interface for quantitative desks that prefer to build their own routing logic. Every Deep Flow score is published as a structured event with versioning, idempotency keys, and replayable history — so a model decision can be reconstructed exactly as it happened, on the day it happened.
- Sub-180ms publish-to-receipt on a co-located WebSocket
- Versioned model output with replayable history per venue
- Sandbox tenant with eighteen months of historical scores
Surface 03 / treasury
Treasury Console — the protocol desk.
A read-and-route console for DAO treasuries and corporate crypto-native balance sheets. Every recommended rebalance comes with a written rationale, a confidence interval, and a one-click refusal that is logged. The console does not custody assets; it tells the operator what to move, and why, in plain English.
- Role-based access with named operators and write-locked approvals
- One-click refusal per recommendation, fully logged
- No custody — treasury keys never leave the operator's stack
current allocation
62% stables · 28% blue-chip LPs · 10% idle
recommended rebalance
rotate 12.4% into USDC / 3pool · arb-leg +0.43%
rationale
venue 0x71f9..3a flagged 0.87 rug probability at T−72h · depth-of-book on blue-chip LP thinning
Chapter III · audited performance
An independent audit found +6.8% APY outperformance — and named its boundaries.
“Across the eighteen-month observation window, the PFN Dai Treasury Console delivered an average APY of +6.8% above the passive DeFi benchmark, net of model versioning and slippage assumptions. The result is reproduced from on-chain evidence and the operator's signed instruction log; it does not extend beyond the observation window, and it is not a forecast.”
We publish the figure because an institutional buyer is entitled to ask for it, and we publish its boundary because no figure deserves to be read without one. The +6.8% result is the audited past, not a forecast. It does not extend past the observation window, it does not adjust for the risk taken to earn it, and it is not a guarantee of any future return.
What we can commit to is the mechanism underneath: the same transformer, the same audit trail, and the same operating controls that produced the figure. Past performance is not indicative of future results. Capital allocation is engineered, not promised.
By the numbers · January 2026
Reference figures, taken from the operating log.
- 184 institutional clients across 27 countries
- 9 of the top 50 crypto-native funds by AUM
- 4.2M on-chain signals per second at peak load
- 180ms end-to-end signal latency across 14 chains
Figures drawn from the operating log reviewed under our SOC 2 Type II control boundary. PFN Dai is an analytics provider, not a broker-dealer.