Review dimensions

Work through all six. For each finding, record severity, location, the problem, the
evidence, and a concrete fix.

1. Claim-evidence integrity

Trace every headline number to its source and confirm it matches (see
claim-evidence-map.md). Specifically:

• The synthetic exactness taxonomy: six applications, max_error 0, 128-bit
  ciphertext sizes 262,282 / 1,310,882 / 2,621,791 bytes, 5× and 10× premiums.
• The real-genome studies (Section 11): E5 mean absolute AF delta ≈ 0.0609; E6 max
  F_ST-like 0.1363 and 624 suppressed rows; E7 recovery rates (1.000 under
  min-N-20, 0.000 under freeze, 0.125 under query budget 5); E8 exact product
  moments and max r2 1.0000.

Flag any value that does not match, any transposed digit, and any table whose rows
disagree with its cited CSV/JSON.

2. Security analysis (Section 9)

• Confirm the coverage matrix (Table 4) and the per-attack catalog agree: every
  threat in the matrix has a structured subsection, and every subsection maps to a
  matrix row.
• Confirm every attack subsection uses the same seven-field template (Adversary,
  Attack, Assets at risk, Platform control, Verification hook, Status, Residual
  risk / evidence) and that each Status is drawn from the fixed vocabulary
  (Prevented / Detectable / Mitigated / Residual / Out of scope), consistently with
  the matrix.

3. Honesty and overclaiming

• The real-genome studies must be framed as aggregate-only workflow demonstrations
  on public reference data — NOT biomedical validation or population-genetics
  estimates.
• The F_ST-like and LD r2 values must be labeled simplified small-panel contrasts;
  the E8 r2 = 1.0000 pair must be flagged as a small-count artifact; the
  below-quorum E8 caveat (25 samples < floor of 30) must be present.
• The certificate honesty box must state that the certificate binds artifacts and
  policy facts and does NOT by itself prove correct computation, correct decryption,
  contributor identity, or bounded output-privacy leakage.
• Flag any sentence that claims more than the evidence supports.

4. Citation and cross-reference integrity

• Every reference R1-R37 is both defined and cited, with no gaps or orphans.
• Every “Section N”, “Table N”, and “Figure N” reference resolves to the right
  target; tables are numbered 1-16 and figures 1-8; the Appendix A index matches the
  body.

5. Reproducibility

• The paper gives a concrete, runnable reproduction path (Appendix B); the commands
  exist; the open CLI, open applications, and public data are sufficient to
  reproduce every result without the hosted service. Optionally run the replicate
  skill and confirm the numbers.

6. Scope and framing

• The contribution claims (Section 2) are supported and appropriately bounded.
• The threat model’s assumptions are stated.
• The limitations are acknowledged: single keyholder, metadata leakage,
  output-privacy as the largest residual risk, and synthetic-plus-small-real
  evidence.