Replicate The Blind Machine experiments
AI agent skillIndependently reproduce the experiments in The Blind Machine paper and confirm its published numbers, using only the open Blind CLI and public data — no hosted service, no credentials, no private records.
Download & use this skill
Download .tar.gz
Install the whole repository into your agent's skills directory in one command — it unpacks as a
ready-to-use replicate/ folder (change -C
to your skills path):
curl -fsSL https://blindmachine.org/skills/replicate/archive | tar xz -C ~/.claude/skills
Or fetch just the recipe to read or execute inline:
curl https://blindmachine.org/skills/replicate.md
# Expected resultsEvery value below is what a correct replication reproduces. The syntheticciphertext sizes are byte-deterministic (a mismatch is a real regression); thereal-genome headline numbers are deterministic given the committed sample andvariant selection rules.## Contents- Synthetic BFV exactness taxonomy (Section 10)- E5 — allele-frequency panel- E6 — cross-population differentiation- E7 — release-policy differencing- E8 — linkage-disequilibrium window## Synthetic BFV exactness taxonomy (paper Section 10)At N=20, L=10, 128-bit security, seed 42. `max_error` must be 0 for every row.| Application | Tier | Ciphertext bytes / contribution | Premium || --- | --- | ---: | ---: || `allele_frequency_count` | additive | 262,282 | 1× || `carrier_count` | additive | 262,282 | 1× || `cohort_histogram` | additive | 262,282 | 1× || `polygenic_score_aggregate` | additive | 262,282 | 1× || `allele_frequency_with_variance` | mult-supporting | 1,310,882 | 5× || `genotype_phenotype_covariance` | mult-supporting | 2,621,791 | 10× |The four additive applications share an identical ciphertext size; the payloadpremium of one multiplicative level is 5× (variance) and 10× (covariance).## E5 — allele-frequency panel- Source: IGSR/1000 Genomes Phase 3 `20130502`, region `22:16050000-17000000`.- 10 public samples, 12 complete-call biallelic SNPs.- Encrypted first moment (`allele_frequency_count`) and first+second moments (`allele_frequency_with_variance`) match the cleartext oracle exactly.- Mean absolute AF deviation vs IGSR global: **0.0609**.- Mean heterozygote rate: **0.3083**. Mean dosage variance: **0.2542**.## E6 — cross-population differentiation- Region `22:16050000-17000000`; 50 samples (10 per super-population), 24 SNPs.- Exact moments vs oracle.- Maximum F_ST-like value **0.1363** at `22:16067208:C:G` (EAS 0.90 vs AMR 0.45).- Panel mean F_ST-like **0.0611**; mean absolute AF deviation vs IGSR global **0.0311**; **624** small-cell group rows suppressed (n < 10).- The F_ST-like statistic is a simplified equal-weight heterozygosity contrast over a small selected panel — not a population-genetics estimate.## E7 — release-policy differencing- Region `22:16050000-17250000`; N=25 vs adjacent N=24, 40 SNPs; the held-out sample carries a nonzero alt dosage at 6 of the 40 positions.- Both encrypted aggregates match the cleartext counts; the exact-count difference recovers the held-out sample's dosage vector exactly.- Recovery rate by policy:| Policy | Exact vector recovery || --- | ---: || No policy / minimum-N = 20 only | 1.000 || Minimum-N = 25 blocks adjacent base | 0.000 || Cohort freeze (single release) | 0.000 || Query budget = 5 | 0.125 || Counts rounded to nearest 5 | 0.850 exact-position, 0.000 exact-vector |- Query-budget curve is linear: recovery = (positions the budget allows) / 40 (budget 1 → 0.025, 5 → 0.125, 10 → 0.250, 20 → 0.500, 40 → 1.000).## E8 — linkage-disequilibrium window (draft `genotype_pair_ld`)- Region `22:16050000-16900000`; 25 samples, 11 adjacent variant pairs.- All five product-moment vectors (`sum_a, sum_b, sum_a2, sum_b2, sum_ab`) match the cleartext oracle exactly.- Maximum `r2` = **1.0000** (pair 4) — a small-count artifact: both variants carry only 3 alternate alleles across 25 samples and perfectly co-occur. This is expected spurious LD, not a population signal.- E8 ran below the application's declared floor of 30 contributors, so it is a compute-path correctness check, not a governed release.Part of the Replicate The Blind Machine experiments agent skill.