Skip to content

Pre-registration: second hold-out, rule 2.19

This file is committed and tagged (prereg-holdout2-v1) before any data of the samples below are collected. It follows the first registration (PREREG_HOLDOUT.md, tag prereg-holdout-v1), whose results are in HOLDOUT_RESULTS.md.

Why a second hold-out

The first hold-out tested rule 2.17. In it the linked withdrawal sender (linked_sender) was again at chance level (Ethereum 14 / 11, R 0.84) and was 11 of the 15 decoy moderate leads. Rule 2.19 therefore makes linked_sender context only. That decision was taken after looking at the first hold-out, so re-scoring the same depositors under 2.19 (moderate, 30 days: 101 / 4, R 0.042, exact interval 0.011–0.111) is exploratory. This registration tests 2.19 on depositors that no earlier analysis has seen.

Frozen rule

  • Code: rule 2.19 as committed with this file. linked_sender is shown as context: no weight, no class, not an independent family (like linked_late since 2.17).
  • Lead sources: early direct link (linked, first withdrawal ≤ 72 h after the voucher's last deposit); labelled shared exchange deposit (shared_deposit); early multi-pool profile (early_profile). strong = leads from two independent sources, moderate = one, weak = otherwise. Amount+timing, gas price, late links and linked senders never raise the class.
  • The files below are not changed until the report is written:
File SHA-256
tornado_demix/heuristics.py 017c372093d20f2caa0dfb091dda268919e12f3024876983ff0e20bd7cc5de71
tornado_demix/demix.py 97fc58e815c3e04588397549f321d0595cb596c50d212a71d8ddfa6ee2948d86
tornado_demix/deposit_addresses.py d65d13ed16be30569b060395aec473a29667aa04e83cad7201bcdb1ad0f60211
tornado_demix/etherscan.py 5f3d6b26aab4f21cb3ad66bdfc89393c2b70c62a731cb843cc18fbe30f31f2c7
  • Attribution labels: unchanged from the first registration (prettydeath/wallet-attribution cca7567; ethereum.csv SHA-256 0392efcc…b81d6, polygon.csv SHA-256 2f311d33…6f07b).

Samples

Eligibility is the same as before (deposits span ≤ 60 days, decoy windows after the pool's first withdrawal, target windows complete before 2026-08-01). No depositor of any earlier sample is eligible.

  1. Ethereum. tools/placebo_eval.py --network ethereum --seed 21 --sample 600 --dir holdout2 --exclude-dir . --exclude-dir post2022 --exclude-dir dar --exclude-dir dar_labels_only --exclude-dir dar_nolabels --exclude-dir holdout1 --exclude-file .cache/labels/placebo/exclude_ens_eval.txt (the ENS list is private; SHA-256 of its sorted, newline-terminated content 26549ae2f6705d00b38ef51147ebe94fa406ee1a5a2e607aa79ee6580869dfc8, unchanged since the first registration). Dry run at tagging time: 55,033 eligible, 1,917 excluded, 600 drawn; sample SHA-256 68e9711c7a0947c409a94ff854e17c65bce3dd35e43516740003f6dd67948347.
  2. Polygon. tools/placebo_eval.py --network polygon --seed 22 --sample 400 --dir polygon2 --exclude-dir polygon1. Dry run at tagging time: 4,444 eligible, 300 excluded, 400 drawn; sample SHA-256 1f6db763abd5de5acfdefa18ea869bcc980ccab795e79cb76e7e98b81a7742e5.

Failed explorer requests are retried; depositors whose runs still fail are listed in the report, never silently dropped or replaced.

Measures

As in the first registration: tools/holdout_report.py per sample, 30-day and 72-hour windows, per signal and per class, R = (L_decoy / N_decoy) / (L_target / N_target), 95 % intervals by bootstrap over depositors (2,000 resamples, seed 20260801) or exact when a side has fewer than five leads; the source combination of every strong candidate.

Hypotheses (rule 2.19)

  • H1. Ethereum, early direct link: upper bound of the R interval (30 days) below 1.
  • H2. Ethereum, count match: the R interval (30 days) contains 1 or lies above it.
  • H3. Ethereum, class moderate: upper bound of R (30 days, the tool's default window) below 0.25.
  • H4. Ethereum, labelled shared deposit: upper bound of R (30 days) below 1 if at least five target leads exist; otherwise too few to judge.
  • H5. Polygon, early direct link: upper bound of R (30 days) below 1.
  • H6. Polygon, class moderate: upper bound of R (30 days) below 0.25.
  • Descriptive only: early profile, linked sender, gas price, strong (counts, and R where the interval rules allow).

A hypothesis that fails is reported as failed. No threshold, window, signal or sample size is changed after collection; any deviation is listed with its reason.