Synthetic BFV Evidence
paper evidence synthetic dataPublic verification package for the paper's six-application BFV experiments.
f77f6eb670b5a0b38fbef977feefefc273baeae853fd7af5435374a1e96dbd63
https://blindmachine.org/verify/paper/bfv-v1
https://blindmachine.org/verify/paper/bfv-v1.json
What This Verifies
This is not a private-cohort computation certificate; it is the public paper reproducibility package.
This package binds the committed reproducer scripts and reference artifacts behind the paper's synthetic BFV claims. It verifies local simulation and benchmark evidence over seeded synthetic cohorts. It is not a ComputationCertificate for a real private cohort run.
6
4
2
33
Claims Checked
These are the machine-independent checks behind the paper's BFV tables and differencing figure.
All six evaluated BFV applications match their plaintext simulation on seeded synthetic inputs.
docs/paper/experiments/verify.py checks exact=true, max_error=0, and feasibility=ok for every generated benchmark cell; the applications array binds the signed bundle digests executed by the harness.
The evaluated catalog separates into four additive-BFV-exact applications and two multiplication-supporting-BFV-exact applications.
docs/paper/experiments/e1_exactness_taxonomy.sh and results/expected/table_b_reference.json bind the six application classes.
At the 128-bit artifact setting, one multiplicative level costs 2.5x the additive tier per encrypted value (80.01 versus 32.02 ciphertext bytes per packed slot, the 5:2 ratio of RNS residues, with the ring degree cancelling). The 5x and 10x per-contributor premiums measured at vector length 10 are a packing artifact of a nearly empty ring: they fall to 2.5x and 5x once the additive ring is filled, and the covariance arm's extra factor is a second uploaded encrypted vector rather than a deeper circuit.
docs/paper/experiments/e3_feasibility.sh sweeps both tiers across the packing boundary (L = 8191, 8192, 16383); verify.py asserts INV-8 (bytes per packed slot equal the closed form 16k, the per-contributor premium flipping 5.0x -> 2.5x between L = 8191 and L = 8192, and covariance costing the same per slot as variance) plus INV-3 (additive < variance < covariance); results/expected/packing_premium_reference.json records the derived per-slot table and results/expected/table_b_reference.json the per-contributor byte counts.
The 128/192/256-bit security matrix remains bit-exact for the six evaluated applications.
docs/paper/experiments/e2_security_matrix.sh and results/expected/security_matrix_reference.json cover the full security-level grid.
K-vs-K+1 differencing recovers one synthetic contributor exactly when cohorts are not frozen.
docs/paper/experiments/e4_differencing.sh writes the differencing raw result; verify.py checks recovered_vector == target_vector.
Reproduce Locally
Run the same local harness. The hosted service is not contacted.
cd docs/paper/experiments && bash run_all.sh
Runs the six-application exactness taxonomy and K-vs-K+1 differencing demonstration.
cd docs/paper/experiments && bash run_all.sh full
Adds the 128/192/256-bit security matrix and feasibility sweep.
python3 docs/paper/experiments/verify.py
Checks already-generated raw outputs and rewrites paper-facing CSV tables.
Signed Application Bundles
These are the content-addressed application payloads executed by the BFV harness.
| Application | Tier | Bundle digest | Env lock |
|---|---|---|---|
| allele_frequency_count | additive-BFV-exact | b94bd9320ea0f15b2ec265ecd0cf855f273548ffb920f395212256f4d4664eed | afd4ed396fee544ee91774f8fe3cc1b9d26d6796558b0fa0897660655785963f |
| carrier_count | additive-BFV-exact | 8691074bf4bbb434cbebb0c61acf6a40cd01987f0a9035a07c4cc484886fb43c | 28a71f683860071c200fd61b679b6a9e2b3ea063ed0b016ea85ac8eb8ab9084a |
| cohort_histogram | additive-BFV-exact | 9f030d1c07b87761e9aed5ca7214a5cac8ed3f18cde2c71d8032bcbfb1a3a953 | 5ad7a840ab78912bfb79e4c41da0556318eabaa76c5fceea7785a5036a7c64c9 |
| polygenic_score_aggregate | additive-BFV-exact | 413d0b16227c861c1b1c24fa6d119013e0a0e3eb58ab573397995076b3b2aa68 | 170d7c7717666bd71adabb3923cd60eafb425a3e2253a18901769795e44179ec |
| allele_frequency_with_variance | mult-supporting-BFV-exact | b48cdffa32c46d2a5de95010ea12e434593b2af2179fcedf7f8e36ebc7245eec | df1559d9c292f359ceb1c0ccb75619688a210e98c4838fdee620dc2ffd048c8d |
| genotype_phenotype_covariance | mult-supporting-BFV-exact | df37c9b8139fda1c3a1e95b5e64f65b6f072dbc22b2545b33516d79908e296de | a61e3c4ac80a0b871ebe5ab113d780210c60db5a0fcf6c0d1796feb1ce38353e |
Paper Tables
The page records the table sources used by the current paper draft.
| Artifact | Rows | Purpose |
|---|---|---|
| docs/paper/experiments/results/table_b_exactness.csv | 6 | Generated source for the BFV exactness taxonomy table. |
| docs/paper/experiments/results/table_c_premium.csv | 3 | Generated source for the additive versus multiplication-supporting payload premium table. |
| docs/paper/experiments/results/security_matrix.csv | 18 | Generated source for the security-level exactness matrix. |
| docs/paper/experiments/results/feasibility.csv | 4 | Generated source for the current feasibility sweep. |
| docs/paper/experiments/results/packing_premium.csv | 9 | Generated source for the per-packed-slot payload premium: both tiers swept across the additive ring boundary (L = 8191, 8192, 16383), with ciphertexts, packed slots, and bytes per packed slot against the 16k closed form. |
Evidence Files
Hashes identify the scripts and reference artifacts for this evidence package.
| Role | Path | SHA-256 | Bytes |
|---|---|---|---|
| source_index | docs/paper/benchmark-results.md | a159bc1dbd855fb12efed1884a766cc406661de617810421b7ab6d62fc7f7863 | 3110 |
| reproducer | docs/paper/experiments/README.md | 05f3316459ee526d980faa1159a52139ccffa070feab6cb0337acbd2d561c500 | 12542 |
| reproducer | docs/paper/experiments/e1_exactness_taxonomy.sh | a3a08c8cf81b4cb3374e570feacac73982693f64338f9b7d0ea068954aca8519 | 806 |
| reproducer | docs/paper/experiments/e2_security_matrix.sh | 0e1c2e513399dce476f05c3c8cbdfef61ec03cc46f27c3b45f1d144ba434dd12 | 770 |
| reproducer | docs/paper/experiments/e3_feasibility.sh | 64a225c0ef78d0b326537d0c8bbcffb73df7ed756ccf15d3cbd7fd0b2d977d3a | 3207 |
| reproducer | docs/paper/experiments/e4_differencing.sh | 1e05db8fbb6c1210f9c8488dcb15f71b8208756e4a7aad27d575748429fa8f9c | 891 |
| reproducer | docs/paper/experiments/lib.sh | 9d8a2f44d56b0a0c0844ba1b044fbe56efc25d5a1d136b375a8e1874de7e2457 | 5347 |
| expected_reference | docs/paper/experiments/results/expected/feasibility_reference.json | 1ff735891b1c41165f42cdd6b55a6130231436d72dbca3856436e5b635410ee3 | 322 |
| expected_reference | docs/paper/experiments/results/expected/packing_premium_reference.json | f7e2077b55a44266efd1d85037af0d7ed1a1414088c2b073997ed81d47871f64 | 9393 |
| expected_reference | docs/paper/experiments/results/expected/security_matrix_reference.json | c30e6701d70b550b5dc1739713cbc56fa1fce1fb2ebd313bbfd85d838771c929 | 2139 |
| expected_reference | docs/paper/experiments/results/expected/table_b_reference.json | b531311d793c0b99606d3b7343c7e29e6ccb9e1b6ea85c86a360a780905d4cfa | 1095 |
| reproducer | docs/paper/experiments/run_all.sh | 24dfd58f62c88683b50ceab6cf91829d5499ee000b23d9a4e97070602a08ff78 | 3909 |
| reproducer | docs/paper/experiments/setup.sh | 796ad5d83118ac6df1641908fdd72f976914b58d0103c0dadd612167abdcd6fd | 3487 |
| reproducer | docs/paper/experiments/verify.py | a46277388301295e02a7c48b6da9c0ac14d7ae2b22036f05960b18fed1a5fbd9 | 35660 |
| benchmark_artifact | docs/paper/artifacts/measure_real_bench.csv | 7a6a391b2e061e78d46d0d1e4b5c94333881326d5b2eab75d33d8dfc2061fe9c | 9573 |
| benchmark_artifact | docs/paper/artifacts/measure_real_bench.json | ea899f8d9e2151dab43592c20d5751925762494cf71cd521036c5cb870234ba1 | 38272 |
| benchmark_artifact | docs/paper/artifacts/measure_real_bench.py | 5cc52bb3d475f4eb6bc6cf8c5be1b1634c5a72829f0b2aa9609a00b5bdb186f1 | 17629 |
| benchmark_artifact | docs/paper/artifacts/measure_run.log | ab47cbaf8581ed9ec275d21ff8300ea81aabe7416094cd793583fa1768cbaeb6 | 6941 |
| benchmark_artifact | docs/paper/artifacts/security_matrix.csv | 0a9319b05ae1a953b1b3727d04df888b42b48cfb73f0c1b68f5564ab9cf98664 | 1285 |
Boundaries
- This package verifies synthetic BFV equality and paper benchmark artifacts, not a real private cohort run.
- The hosted service is not contacted by the reproduction commands.
- Generated raw JSON and CSV outputs are intentionally excluded because they are mutable outputs of the reproducer.
- The current package is draft metadata; the source commit should be replaced with a release commit or tag before camera-ready submission.
- Timing, CPU, memory, and cost are hardware-dependent; exactness and invariant pass/fail are the portable checks.