Independently review The Blind Machine paper

AI agent skill

Independently review The Blind Machine paper the way a careful, skeptical referee would — verifying every claim against evidence rather than accepting it.

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 review/ folder (change -C to your skills path):

curl -fsSL https://blindmachine.org/skills/review/archive | tar xz -C ~/.claude/skills

Or fetch just the recipe to read or execute inline:

curl https://blindmachine.org/skills/review.md

Claim-evidence map

The headline numbers in the paper and where each one comes from. Use this to
fact-check dimensions 1 and 3. Every value here is what a correct paper states; a
mismatch in the manuscript is a finding.

Synthetic exactness (Section 10)

Claim	Value	Source
Additive ciphertext / contribution (128-bit)	262,282 bytes	`experiments/results/table_b_exactness.csv`
Variance ciphertext (premium)	1,310,882 bytes (5×)	same
Covariance ciphertext (premium)	2,621,791 bytes (10×)	same
Exactness	`max_error == 0` for all six apps	`verify.py` invariants

E5 — allele-frequency panel

Claim	Value	Source
Mean absolute AF delta vs IGSR global	0.0609	`real_human_dna_igsr_2026_07_09/results/report.md`
Mean heterozygote rate	0.3083	same / `genotype_distribution.csv`
Mean dosage variance	0.2542	same
Encrypted vs oracle	exact (counts + first/second moments)	`blindmachine_results.json`

E6 — cross-population differentiation

Claim	Value	Source
Max F_ST-like	0.1363 at `22:16067208:C:G`	`public_af_fst_2026_07_09/results/fst_summary.csv`
Panel mean F_ST-like	0.0611	`report.md`
Mean absolute AF delta vs IGSR global	0.0311	`report.md`
Suppressed small-cell group rows (n < 10)	624	`report.md`
F_ST-like definition	simplified equal-weight heterozygosity contrast, not the Weir-Cockerham estimator	Section 11.2

E7 — release-policy differencing

Claim	Value	Source
Cohorts	N=25 vs adjacent N=24, 40 SNPs	`beacon_release_policy_2026_07_09/results/policy_risk_summary.csv`
Recovery, minimum-N = 20 only	1.000	same
Recovery, cohort freeze / min-N-25-blocks	0.000	same
Recovery, query budget = 5	0.125	`query_budget_curve.csv`
Query-budget curve	linear: recovery = probed positions / 40	`query_budget_curve.csv`

E8 — linkage-disequilibrium window (draft)

Claim	Value	Source
Encrypted product moments	exact vs oracle (all 5 vectors)	`public_ld_window_2026_07_09/results/{ld_pairs.csv,cleartext_oracle.json}`
Max `r2`	1.0000 (pair 4) — small-count artifact	`ld_pairs.csv`
Quorum	ran at N=25, below the declared floor of 30	manifest / Section 11.4

Data provenance (all four studies)

Source: IGSR/1000 Genomes Project Phase 3 integrated callset 20130502, GRCh37,
chromosome 22 (references R35, R36).
Public, openly consented reference data; aggregate-only committed outputs.

Part of the Independently review The Blind Machine paper agent skill.