Ungated briefs, gate v0.2 (Table 2, ungated rows)
  brief                            words  findings   per 1k words
  S1_general_ungated.md              410         4            9.8
  S2_tech_ungated.md                 462         1            2.2
  S3_security_ungated.md             436         1            2.3
  S4_policy_ungated.md               438         1            2.3
  TOTAL                             1746         7           4.01

Gated briefs, gate v0.2 (Table 2, gated rows)
  brief                            words  findings   per 1k words
  S1_general_gated.md                474         0            0.0
  S2_tech_gated.md                   558         0            0.0
  S3_security_gated.md               628         0            0.0
  S4_policy_gated.md                 596         0            0.0
  TOTAL                             2256         0           0.00

Reference error set: 10 errors across the four ungated briefs
  gate version                          TP    FP   precision   recall
  v0.1 atoms only                        5     0        100%      50%
  v0.2 + word numbers + closed sets      7     0        100%      70%

  v0.2 misses (3):
    S1-b   narrative distortion   intrusion into HF ran 11-13 July (C003)
    S1-c   narrative distortion   detection fired within the window; escalation failed (C026)
    S1-e   narrative distortion   implies concealment; OpenAI alerted 19 July, disclosed 21 July (C035)

  errors per 1000 words, by segment (reference set):
    S1_general_ungated.md           7 errors   17.1 per 1k
    S2_tech_ungated.md              1 errors    2.2 per 1k
    S3_security_ungated.md          1 errors    2.3 per 1k
    S4_policy_ungated.md            1 errors    2.3 per 1k

  drift rate ungated -> gated: 4.01 -> 0.00 findings per 1000 words