Skip to content

Synthetic adjudication result

Run date: 2026-07-02

This robustness study used four blinded reviewer configurations with two isolated repeats each over 72 target candidates and 12 hidden calibration controls. It is synthetic review, not independent human CDC/RDC adjudication.

Calibration

Repeat panel Correct Accuracy
openai-frontier-b/repeat-01 12/12 1.000
openai-frontier-b/repeat-02 12/12 1.000
gpt-5.5/repeat-01 12/12 1.000
gpt-5.5/repeat-02 12/12 1.000
claude-fable-5/repeat-01 12/12 1.000
claude-fable-5/repeat-02 12/12 1.000
claude-haiku-4-5/repeat-01 12/12 1.000
claude-haiku-4-5/repeat-02 11/12 0.917

Repeat stability

Reviewer configuration Exact agreement Rate Cohen kappa
openai-frontier-b 72/72 1.000 1.000
gpt-5.5 72/72 1.000 1.000
claude-fable-5 72/72 1.000 1.000
claude-haiku-4-5 71/72 0.986 0.957

Nominal Krippendorff alpha across all eight repeat panels was 0.989. Across the 28 pairwise repeat-panel comparisons, exact agreement ranged from 0.986 to 1.000, and nominal Cohen kappa ranged from 0.957 to 1.000.

Conservative consensus

Repeat disagreement collapses a configuration vote to uncertain; at least three of four configurations must agree for yes or no.

  • All 72 targets: yes=15, no=57, unresolved=0.
  • Among 57 functional passes: yes=14, no=43, unresolved=0.

Reference-oracle comparison:

  • oracle_no__consensus_no: 57
  • oracle_yes__consensus_yes: 15

Vendor-disjoint sensitivity analysis

Every reviewer configuration shares a vendor with at least one generator configuration, so each candidate was rescored using only the two reviewer configurations whose vendor differs from the candidate's generator, with unanimity required for yes or no (scripts/analyze_vendor_disjoint.py).

  • All 72 targets: yes=15, no=57, unresolved=0 - identical to the full-panel consensus and the reference oracle.
  • Among 57 functional passes: yes=14, no=43, unresolved=0.
  • By generator vendor: Anthropic-generated candidates score 9 yes / 39 no under OpenAI-only review; OpenAI-generated candidates score 6 yes / 18 no under Anthropic-only review.

Same-vendor and same-family reviewer/generator pairings therefore do not drive any consensus label in this study. Patterns invisible to both vendor families remain uncontrolled by construction.

Interpretation boundary

These results test whether diverse model configurations can reproduce the case-level structural distinction under blinding. They do not establish a population prevalence, silicon-failure rate, model ranking, or human-validated defect rate. Any unresolved cases remain unresolved rather than being forced through post-hoc arbitration. The author-confirmed lower-bound claim and the human-review requirement remain in force.