Security notes — allele_frequency_count

Scoped to this bundle. The platform-wide threat model lives in
docs/manifesto.md, docs/requirements.md, and docs/simulation_mode.md §5.
Kerckhoffs applied to a product: no guarantee rests on the secrecy — or the
honesty — of the server. Don’t trust, verify.

Trust classes (what may cross the wire)

Only ENCRYPTED and PUBLIC are ever uploaded. 00_keygen.py writes the secret key
to secret_context.tenseal, which is used only by 40_decrypt.py on the
researcher’s machine. There is no /api/v1 endpoint that accepts a secret key.

Server holds no secret key

30_compute_encrypted.py — the only server-side stage, a kit shim that runs
server.py’s compute — loads the public context plus ciphertexts and
homomorphically adds. It defensively refuses a context that carries a secret key
(context.is_private() → error). The server
therefore never sees a single plaintext genotype; it operates on ciphertext and
returns ciphertext. Decryption happens only where the secret key lives: locally.

The append-1 sentinel is NOT a MAC

The trailing sentinel slot decrypts to the exact contributor count N, and
dropping one upload yields N−1 (test: test_sentinel_tracks_dropped_upload). It
catches mechanical corruption / miscounting — it gives no guarantee that
contributions are distinct, genuine, or non-Sybil. Call it what it is: an
integrity check, not authenticity.

What FHE here does and does not hide

• Hides: individual genotype vectors from the server (inputs are ciphertext).
• Does not hide: the released aggregate itself, and metadata (researcher
  identity, participant count/timing, ciphertext sizes, protocol choice).
• Differencing (K vs K+1): the statistic leaks an individual if you can
  compute A_{K+1} − A_K. aggregate_only + min_contributors ≥ 20 +
  allowed_runs_per_project: 1 (cohort freeze + min-N + run cap) mitigate
  this; they are not a complete defense. Overlapping/Sybil differencing across
  separately frozen cohorts needs DP + cross-job query budgets (v2). Documented,
  not hand-waved — see docs/simulation_mode.md §5.
• Verify-by-re-execution is determinism, not zero-knowledge. Re-running
  30_compute_encrypted.py on the same ciphertexts reproduces a bit-identical
  result digest; it proves the computation, it is not a ZK proof.

Exactness / parameter safety

BFV is exact in Z_t. The plaintext modulus must satisfy t > max coordinate sum = 2·N. The default t = 1032193 (a 20-bit batching prime) stays exact for N
up to ~500k; a real run at larger N must raise t (or the simulation feasibility
sweep will report infeasible-at-these-params on overflow). The sentinel sum is
N, always ≪ t.