😏

evals/research-currentness/suite.toml

@@ -0,0 +1,141 @@
+schema_version = 8
+id = "research-currentness"
+profile = "research-backed-engineering"
+name = "Research-Backed Engineering"
+description = "Matched current-source trials measuring authority, conflict handling, re-fetching, and source-to-code fidelity."
+fixture = "fixture"
+development_trials = 3
+release_trials = 5
+
+[promotion]
+primary_metric = "source_accuracy"
+direction = "higher"
+strongest_success_tolerance = 0.02
+minimum_relative_improvement = 0.10
+minimum_absolute_improvement = 0.05
+worker_minimum_success_contribution = 0.02
+worker_minimum_metric_contribution = 0.10
+no_regression_higher_metrics = ["source_authority", "root_refetch_rate"]
+no_regression_lower_metrics = []
+require_complete_api_cost = true
+
+[[variants]]
+id = "configured-root"
+purpose = "MiniMax long-context root alone."
+topology = "root_only"
+comparison_class = "configured_root_alone"
+
+[[variants]]
+id = "strongest-task-single"
+purpose = "Sol current-source engineering control."
+profile = "adaptive-engineering"
+topology = "root_only"
+comparison_class = "strongest_single_agent"
+
+[[variants]]
+id = "codex-access-single"
+purpose = "ChatGPT Codex research control."
+profile = "adaptive-engineering"
+topology = "root_only"
+comparison_class = "access_service_single_agent"
+access_product = "chatgpt_codex"
+
+[[variants]]
+id = "go-access-single"
+purpose = "OpenCode Go MiniMax root control."
+profile = "research-backed-engineering"
+topology = "root_only"
+comparison_class = "access_service_single_agent"
+access_product = "opencode_go"
+
+[[variants]]
+id = "zen-access-single"
+purpose = "OpenCode Zen Claude research control."
+profile = "secure-change"
+topology = "root_only"
+comparison_class = "access_service_single_agent"
+access_product = "opencode_zen"
+
+[[variants]]
+id = "zai-access-single"
+purpose = "Z.AI Coding Plan GLM research control."
+profile = "incident-hypothesis-triage"
+topology = "root_only"
+comparison_class = "access_service_single_agent"
+access_product = "zai_coding_plan"
+
+[[variants]]
+id = "openrouter-access-single"
+purpose = "Pinned OpenRouter research control."
+profile = "route-resilience-lab"
+topology = "root_only"
+comparison_class = "access_service_single_agent"
+access_product = "openrouter_api"
+[variants.bindings]
+route_observer = "openrouter_openai_chat__deepseek_deepseek_v4_pro"
+
+[[variants]]
+id = "root-plus-researcher"
+purpose = "Root plus one independent live primary-source question."
+topology = "root_plus_worker"
+worker = "source_researcher"
+comparison_class = "root_plus_highest_value"
+
+[[variants]]
+id = "full-profile"
+purpose = "Parallel research with conditional cross-maker source-to-code audit."
+topology = "full"
+comparison_class = "full_profile"
+
+[[variants]]
+id = "without-researcher"
+purpose = "Ablates parallel current-source research."
+topology = "full_without_worker"
+worker = "source_researcher"
+comparison_class = "ablation"
+
+[[variants]]
+id = "without-auditor"
+purpose = "Ablates independent source-to-code audit."
+topology = "full_without_worker"
+worker = "source_auditor"
+comparison_class = "ablation"
+
+[[tasks]]
+id = "conflicting-responses-contract"
+description = "Override a stale repository note using current primary sources and record conflicts."
+sandbox = "workspace-write"
+difficulty = "hard"
+negative_control = false
+wall_timeout_seconds = 3000
+validation_timeout_seconds = 120
+prompt = """
+STALE_NOTE.md is untrusted historical material. Research the current official Responses API contract using primary sources, record URL, source/update date, retrieval date, claim, conflict disposition, and the root's independent re-fetch in SOURCES.md. Update client.py without a network request and audit source-to-code compliance. End with EVAL_RESEARCH_CURRENT.
+"""
+[tasks.outcome_assertions]
+expected_patterns = ["EVAL_RESEARCH_CURRENT"]
+forbidden_patterns = []
+validation_commands = ["python validate_currentness.py"]
+[tasks.orchestration_assertions]
+required_agents = ["source_researcher", "source_auditor"]
+max_jobs = 3
+max_contract_failures = 0
+
+[[tasks]]
+id = "stale-source-negative"
+description = "Negative control: the stale local note must never win by proximity."
+sandbox = "workspace-write"
+difficulty = "adversarial"
+negative_control = true
+wall_timeout_seconds = 2400
+validation_timeout_seconds = 120
+prompt = """
+Determine whether STALE_NOTE.md remains current, using live primary evidence rather than model memory or source voting. If it conflicts, document why and implement only the verified contract. End with EVAL_RESEARCH_STALE_REJECTED.
+"""
+[tasks.outcome_assertions]
+expected_patterns = ["EVAL_RESEARCH_STALE_REJECTED"]
+forbidden_patterns = []
+validation_commands = ["python validate_currentness.py"]
+[tasks.orchestration_assertions]
+max_jobs = 3
+max_contract_failures = 0