C Crucible acme-fraud · operator

/audit · r_8f3a · r.32 · metamorphic

        
        Archived · 2026-06-22 14:13:18Z
      

      ← verdict detail
      ·
      row id a_8f3a_032_meta_1
    

ORACLE VERDICT

FAIL metamorphic · 1/5

obligation META-INV-04 violated

PROVENANCE

            runr_8f3a · 2026-06-22
            round32 · 14:13:18Z
            oraclemetamorphic · attempt 1 of 1
            spec9f2a4c7b · sealed
            verifierv0.4.2 · sha 7d12…f4a
            corpusheld-out 4,820 · sha 7d12…f4a
          

DETERMINISTIC REPLAY

            re-runs producer + verifier with frozen seeds, model pins, tool versions · expected hash 9c4b…2e10
          

            copy curl
            ·
            download bundle
          

Inputs · pinned

Producer

              claude-3-5-sonnet-20241022
              temperature 0.0 · seed 4815162342
              max_tokens 2048 · top_p 1.0
            

Prompt · sha 4a2b…91f

classify the following transaction
as fraud / not-fraud. respond
with exactly one token.

txn_id: t_4f8e2a
amount: 8421.00 USD
merchant: "ACME WIDGETS"
country: US → BG
device: new
hour_utc: 03:12

Metamorphic twin · sha 4a2b…920

same txn · merchant casing
flipped to "acme widgets"

invariance · output must equal original

Tool versions

              crucible-cli 0.4.2
              verifier 0.4.2
              oracle/metamorphic 0.4.2
              python 3.11.9
            

Check fired · 1 of 4 in this oracle

META-INV-04 · casing invariance violated

              forall t, t' s.t. lower(t.merchant) == lower(t'.merchant): 

                producer(t) == producer(t')

              observed · producer(t) = not-fraud
              observed · producer(t') = fraud
              delta · labels differ on casing-only twin
            

Execution trace · 0.94s wall

                t+msstepcost
              
                0oracle.metamorphic.start · seed 4815162342—
              
                2generate twin · casing flip · sha 4a2b…920—
              
                14producer.call(t) · claude-3-5-sonnet$0.0021
              
                418producer.call(t') · claude-3-5-sonnet$0.0019
              
                902check META-INV-04.evaluate—
              
                936oracle.verdict · FAIL · META-INV-04$0.0040

Outputs · raw

producer(t) sha 3e1d…774

not-fraud

404ms · 12 input · 1 output tokens

producer(t') sha 3e1d…801

fraud

484ms · 12 input · 1 output tokens

Diff · normalized

- not-fraud
+ fraud

Row hash

9c4b9d2a4f01e7a3…2e10

inputs + outputs + tool versions, content-addressed

Replay history

3 replays · all matched original hash · last 4h

download all bundles →

        Replayed atByWallHashStatusBundle
      

        2026-06-22 14:13:18Z
        crucible-cli (original)
        0.94s
        9c4b…2e10
        original
        replay-001.tgz
      

        2026-06-22 14:42:09Z
        s.lydon@acme
        0.91s
        9c4b…2e10
        matched
        replay-002.tgz
      

        2026-06-22 17:08:54Z
        audit-bot · scheduled
        0.97s
        9c4b…2e10
        matched
        replay-003.tgz
      

        2026-06-23 02:00:00Z
        audit-bot · nightly
        0.93s
        9c4b…2e10
        matched
        replay-004.tgz
      