Topic 41 notes — On-chain & crypto analytics

Predictions vs measurements

Two mechanics worth memorizing.

FIFO is lossless, and that is the whole argument. Haircut and FIFO both conserve the stolen total exactly — neither invents money. The difference is that haircut turns “stolen” into a real number between 0 and 1 that gets multiplied at every merge, so after two hops nothing can be reversed and everybody holds a trace. FIFO keeps a satoshi stolen or not-stolen, so provenance survives arbitrarily many hops and can be traced backwards. Same conservation law, 114× narrower answer (32 UTXOs vs 3657), and the real chain agrees: Linode taints 1.35% of addresses under FIFO and 93% under haircut.

A false merge is permanent; a missed merge is not. Co-spending is a property of the protocol — you cannot be co-spent with someone whose key you do not hold — so Heuristic 1 has precision 1.000 by construction and keeps it at every parameter setting. The change heuristic keys on a habit, so it can be wrong, and because union-find is transitive one error fuses two whole components. That asymmetry is why recall 0.041 at precision 1.000 can be worth more than recall 0.397 at precision 0.089.

Guide-question checklist

• reading-fistful-of-bitcoins.md Q1–Q5
• reading-bitcoin-redux.md Q1–Q5
• reading-blocksci.md Q1–Q5
• reading-elliptic-aml.md Q1–Q5

Paper numbers worth keeping

Cross-topic threads (worked)

• Topic 39 ↔ 41: address clustering is entity resolution. Same union-find, same pair precision/recall scorer, same “one false merge is permanent” hazard. The difference is that Fellegi–Sunter learns its weights from unlabelled data with EM while Definition 4.3 is four hand-written conditions — and lane 3 measures exactly what the hand-written version costs when its assumption (fresh change address) is violated 5% of the time.
• Topic 40 ↔ 41: both score a graph an adversary reads. FRAUDAR (39) answered camouflage by picking a metric the fraudster cannot move; here the equivalent move is preferring co-spending (protocol) to change detection (habit), and FIFO (lossless) to haircut (diffusing). Elliptic shows what happens when there is no such move available: the dark market shuts and every model dies.
• Topic 12 ↔ 41: BlockSci’s inline layout with 19% duplication, 32-bit ids and 60-bit packed values is the columnar argument. Its 500× win over graph databases on calculate fee is a scan win, not a traversal win — read Table 3 by row and the losses are all full scans with arithmetic.
• Topic 32 ↔ 41: which makes Table 3 an HTAP brief. A graph engine keeps the point lookups (Neo4j-with-index beats single-threaded BlockSci on Tx locktime > 0) and needs a columnar side-store for the scans. That is M41’s third deliverable.
• Topic 1 ↔ 41: the three taint policies are a RUM triangle. Poison is O(1) state and useless; haircut is one float per output and exact in total, useless in detail; FIFO is a queue per UTXO — the only one that answers the question, and the only one with an unbounded space term (hence reduce_taint).
• Topic 33 ↔ 41: Elliptic’s 49 time steps have no edges between them, which deletes time-respecting paths by construction. EvolveGCN’s small gain over GCN is the direct consequence.
• Topic 25 ↔ 41: the cautionary result. A hand-built one-hop aggregation fed to Random Forest beat a learned two-layer GCN on the same graph. Build the aggregates and measure before reaching for a GNN — and note that the best configuration used the GCN as a feature extractor rather than a classifier.
• Topic 36 ↔ 41: BlockSci’s “infinite COST” conjecture rests on blockchain data being graph-structured and therefore hard to partition. That is topic 36’s vertex-cut problem, used as an argument for scaling vertically instead.

Open questions

• reduce_taint bounds queue fragmentation by coalescing adjacent same-name runs, but a chain that alternates provenance at every hop defeats it. What is the actual worst case on a real chain, and does it make FIFO taint storable as a graph property or not? (Exercise 3.)
• Meiklejohn buys precision with latency (13% → 0.17% by waiting a week). Is there an online version — a confidence that decays as evidence arrives — or is the delay fundamental? (Exercise 5.)
• The Elliptic labels were produced partly by address clustering, and lane 3 shows clustering can collapse. Nobody in the literature seems to have measured how much label noise that introduces.
• Bitcoin Redux’s own conclusion is that most victim funds never touched the chain at all (hosted wallets, off-chain settlement). What fraction of on-chain analysis conclusions survive that, and is there a published estimate?