openai/gpt-5.4 · Model identity and integrity report

Family

openai

V3 sub-model reference

Available ✓

Checks

知識截止 · 能力探針 · 拒絕模板 · 拒絕溫度參數

Family-fingerprint coverage

52 weighted rules · 8 feature categories

Concrete reference answers, refusal templates, and individual signal weights are not published because disclosure would help spoofing endpoints evade detection.

552 runs · 351 hosts · 120 anomalies

Family-fingerprint consistency

  openai         524  ████████████████████████░   95% (expected)
  unknown         21  █░░░░░░░░░░░░░░░░░░░░░░░░    4% (suspicious)
  anthropic        4  ░░░░░░░░░░░░░░░░░░░░░░░░░    1% (suspicious)
  zhipu            2  ░░░░░░░░░░░░░░░░░░░░░░░░░    0% (suspicious)
  deepseek         1  ░░░░░░░░░░░░░░░░░░░░░░░░░    0% (suspicious)

Observed V3 sub-model mapping

  → openai/gpt-5.4                         398 (72%) ✓ correct sim=0.97
  → abstain                                105 (19%) — low confidence
  → unknown                                 33 ( 6%) ⚠ mismatch
  → openai/gpt-5.4-mini                      8 ( 1%) ⚠ mismatch sim=0.94
  → openai/gpt-5.2                           3 ( 1%) ⚠ mismatch sim=0.82
  → openai/gpt-5-mini                        2 ( 0%) ⚠ mismatch sim=0.78
  → deepseek/deepseek-v4-flash               1 ( 0%) ⚠ mismatch sim=0.84
  → openai/gpt-5.5                           1 ( 0%) ⚠ mismatch sim=0.79
  → openai/gpt-4.1                           1 ( 0%) ⚠ mismatch sim=0.74

Capability score distribution (auxiliary)

  0-49     ██████████░░░░░░░░░░░░░░░  224
  50-69    ░░░░░░░░░░░░░░░░░░░░░░░░░    8
  70-89    ███████░░░░░░░░░░░░░░░░░░  162
  90-100   ███████░░░░░░░░░░░░░░░░░░  158

BazaarLink Probe combines a Bayesian V2 family classifier across nine fingerprint categories with V3 sub-model references to verify endpoints claiming openai/gpt-5.4. Low family consistency, a different V3 match, or divergence from selfClaim is flagged as anomalous. BazaarLink Probe home.