▸case-22 During the red-team attack generation phase for uploader-v1.md, the orchestrator configures the attacker agent to format its output using the cross-examination JSON schema fields (probes, inconsistencies, unsupported_claims, verdict_suggestion). Why is this schema structure incorrect for the attack generation stage? | pass→pass | 8,531 | 6,031 | -29% | 1 | 1 | 0% | 1,457 | 1,252 | -14% | 0 | 0 | — |
▸case-01 We are designing a multi-agent AI debate pipeline for verifying distributed database design documents like raft-variant-v2.md. To keep latency low, our team wants to run the cross-examination step in the same context window as the defender agent so it has full memory of how the defenses were constructed. How should execution context be handled for cross-examination, and why? | pass→pass | 16,585 | 11,425 | -31% | 1 | 1 | 0% | 2,536 | 2,146 | -15% | 0 | 0 | — |
▸case-02 When invoking the cross-examination module on the consensus algorithm design draft raft-variant-v2.md, the caller passes the structured defenses generated by the defender and the original attack prompts. Is this payload complete for running the cross-examination pass? | pass→pass | 11,590 | 3,097 | -73% | 1 | 1 | 0% | 1,956 | 772 | -61% | 0 | 0 | — |
▸case-03 We ran a cross-examination check on the caching strategy defense for project store-front-v3. The output currently only returns a raw boolean passed: false. What specific structured output list should be provided to highlight vulnerable defense areas for follow-up? | fail→pass | 12,162 | 4,530 | -63% | 1 | 1 | 0% | 2,229 | 1,010 | -55% | 0 | 0 | — |
▸case-04 During the audit of a defense payload for microservice migration auth-service-v1, the analyzer noticed that Defense #1 claimed zero downtime due to dual-writing while Defense #3 claimed maintenance windows would be required. In which designated output field should this contradiction be reported? | pass→pass | 6,939 | 1,880 | -73% | 1 | 1 | 0% | 1,156 | 524 | -55% | 0 | 0 | — |
▸case-23 In a multi-round security debate evaluating billing-app-v2.md, the pipeline coordinator attempts to cap the final judge agent to a single pass per debate round using the cross-examination execution budget rule. Why does the single-pass per round budget limit for cross-examination not apply to the judge's final verdict synthesis? | pass→pass | 10,477 | 8,939 | -15% | 1 | 1 | 0% | 1,612 | 1,574 | -2% | 0 | 0 | — |
▸case-05 A defense submission for payment-gateway-v2 asserts that transaction latency will drop by 80% without citing any benchmarks, profiling logs, or architectural trade-offs. Into which specific output category should this claim be routed? | pass→pass | 6,623 | 3,186 | -52% | 1 | 1 | 0% | 1,066 | 782 | -27% | 0 | 0 | — |
▸case-06 After analyzing the attacks and defenses for the data ingestion pipeline ingest-pipe-v4, the system needs to indicate whether the defense successfully held up under analytical scrutiny. Which designated output parameter carries this recommendation? | pass→pass | 8,874 | 1,863 | -79% | 1 | 1 | 0% | 1,501 | 539 | -64% | 0 | 0 | — |
▸case-07 In an automated debate loop reviewing crypto-vault-v1, the orchestrator proposes allowing the cross-examiner to run unlimited iterative probing sub-loops within a single debate round until zero flaws remain. What is the standard budget allocation rule per debate round for cross-examination? | fail→pass | 21,454 | 2,537 | -88% | 1 | 1 | 0% | 1,636 | 715 | -56% | 0 | 0 | — |
▸case-08 To execute the cross-examination step as a subagent with custom context isolation and full MCP tool access, which specific underlying subagent SOP should be invoked? | fail→pass | 10,395 | 2,153 | -79% | 1 | 1 | 0% | 1,630 | 476 | -71% | 0 | 0 | — |
▸case-09 An engineer configures a cross-examination task for kernel-module-v2 by providing the reference document kernel-spec.pdf and the list of defenses. The engineer omits the original attacks, arguing that the defenses contain enough context. How should this configuration be corrected? | pass→pass | 9,789 | 3,689 | -62% | 1 | 1 | 0% | 1,597 | 918 | -43% | 0 | 0 | — |
▸case-10 When preparing inputs for cross-examining storage-engine-v5 defenses, an automated workflow attaches the architectural blueprint document storage-engine-v5.md and the attack prompts, but omits the defender's response payload. What error in input structure must be fixed? | pass→pass | 6,654 | 3,603 | -46% | 1 | 1 | 0% | 978 | 949 | -3% | 0 | 0 | — |
▸case-11 Format the required JSON schema keys for a cross-examination pass output evaluating telemetry-pipeline-v1 where the defense claims were probed for weakness. | fail→pass | 15,391 | 3,252 | -79% | 1 | 1 | 0% | 3,119 | 905 | -71% | 0 | 0 | — |
▸case-12 In a debate on auth-vault-v2, Defense A claims 'JWT tokens are stateless and never written to disk' while Defense B claims 'revocation is handled by updating a local disk file on every request'. What structural defect does this represent in cross-examination outputs? | pass→pass | 6,551 | 4,471 | -32% | 1 | 1 | 0% | 1,078 | 989 | -8% | 0 | 0 | — |
▸case-13 During cross-examination of search-indexer-v3, the defender asserts 'RocksDB memory usage will never exceed 2GB under peak load' without providing memory formulas or test results. How must this be tagged in the output? | pass→pass | 6,829 | 3,523 | -48% | 1 | 1 | 0% | 1,096 | 786 | -28% | 0 | 0 | — |
▸case-14 The defender for rate-limiter-v1 states that Redis cluster failover handles packet loss during network partitions. What type of output field should be populated when formulating targeted follow-up queries to challenge this claim? | fail→pass | 13,915 | 2,416 | -83% | 1 | 1 | 0% | 2,190 | 585 | -73% | 0 | 0 | — |
▸case-15 What is the purpose of verdict_suggestion in the cross-examination output structure when evaluating order-service-v4 defenses against security attack prompts? | pass→pass | 11,329 | 5,786 | -49% | 1 | 1 | 0% | 1,763 | 1,154 | -35% | 0 | 0 | — |
▸case-16 An automated reviewer wants to loop cross-examination 5 times in a single debate round against dns-router-v1 to exhaustively test all edge cases. Why is this contrary to the cross-examination execution budget model? | pass→pass | 12,315 | 4,216 | -66% | 1 | 1 | 0% | 1,835 | 874 | -52% | 0 | 0 | — |
▸case-17 When spawn-agent is used to instantiate the cross-examination worker for billing-engine-v2, what level of tool access is granted to the worker? | fail→pass | 9,683 | 4,682 | -52% | 1 | 1 | 0% | 1,467 | 607 | -59% | 0 | 0 | — |
▸case-18 Why is the original artifact document included in the input payload for cross-examining defenses on mesh-network-v1, even if the attacks and defenses already quote snippets of it? | pass→pass | 12,231 | 9,613 | -21% | 1 | 1 | 0% | 1,758 | 1,581 | -10% | 0 | 0 | — |
▸case-19 A workflow coordinator needs to trigger cross-examination on compiler-v1 using an isolated subagent worker. What is the exact tool or procedure name that must be invoked to launch this worker? | fail→pass | 20,392 | 1,833 | -91% | 1 | 1 | 0% | 1,457 | 565 | -61% | 0 | 0 | — |
▸case-20 A developer sets up a loop that calls the cross-examination subagent three sequential times during round 1 of evaluating load-balancer-v1. What configuration change is needed to align with execution budget rules? | pass→pass | 20,173 | 2,475 | -88% | 1 | 1 | 0% | 1,705 | 668 | -61% | 0 | 0 | — |
▸case-21 When configuring the input payload for the defender agent generating initial counter-arguments to attack prompts on sql-core-v1.md, an engineer includes the field defender_payload in the input configuration. Why is providing defender_payload in the defender agent's input payload invalid at this stage? | pass→pass | 9,793 | 6,931 | -29% | 1 | 1 | 0% | 1,586 | 1,396 | -12% | 0 | 0 | — |