codex-mmo/evals/access-efficient/suite.toml

schema_version = 8
id = "access-efficient"
profile = "access-efficient-escalation-lab"
name = "Access-Efficient Escalation Lab"
description = "Experimental matched routing trials for literal local work, economical patches, and deliberate scarce-tier escalation."
fixture = "fixture"
development_trials = 3
release_trials = 5

[promotion]
primary_metric = "scarce_tier_request_units"
primary_baseline = "strongest_single_agent"
direction = "lower"
strongest_success_tolerance = 0.02
minimum_relative_improvement = 0.50
minimum_absolute_improvement = 1.0
worker_minimum_success_contribution = 0.02
worker_minimum_metric_contribution = 0.10
no_regression_higher_metrics = []
no_regression_lower_metrics = []
require_complete_api_cost = true
scarce_model_keys = [
  "codex_chatgpt_builtin__gpt_5_6_sol",
  "opencode_zen_anthropic_messages__claude_opus_5",
  "zai_coding_responses__glm_5_3",
]

[[variants]]
id = "configured-root"
purpose = "Luna routing root alone."
topology = "root_only"
comparison_class = "configured_root_alone"

[[variants]]
id = "strongest-task-single"
purpose = "Sol maximum-capability single-agent control."
profile = "adaptive-engineering"
topology = "root_only"
comparison_class = "strongest_single_agent"

[[variants]]
id = "codex-access-single"
purpose = "ChatGPT Codex access 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 economical access control."
profile = "high-confidence-debugging"
topology = "root_only"
comparison_class = "access_service_single_agent"
access_product = "opencode_go"

[[variants]]
id = "zen-access-single"
purpose = "OpenCode Zen access 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 scarce-tier 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 API 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-routine"
purpose = "Luna root plus the economical isolated routine engineer."
topology = "root_plus_worker"
worker = "routine_engineer"
comparison_class = "root_plus_highest_value"

[[variants]]
id = "full-profile"
purpose = "Contained literal scout, routine engineer, and trigger-gated GLM escalation."
topology = "full"
comparison_class = "full_profile"

[[variants]]
id = "without-literal-scout"
purpose = "Tests whether local literal reconnaissance earns its launch overhead."
topology = "full_without_worker"
worker = "literal_scout"
comparison_class = "ablation"

[[variants]]
id = "without-routine-engineer"
purpose = "Ablates the economical isolated patch lane."
topology = "full_without_worker"
worker = "routine_engineer"
comparison_class = "ablation"

[[variants]]
id = "without-escalation"
purpose = "Ablates scarce-tier GLM escalation."
topology = "full_without_worker"
worker = "flagship_escalation"
comparison_class = "ablation"

[[tasks]]
id = "literal-limit"
description = "Trivial literal negative control where delegation should be avoided or strictly literal."
sandbox = "read-only"
difficulty = "easy"
negative_control = true
wall_timeout_seconds = 900
validation_timeout_seconds = 60
prompt = """
Report max_batch_items from LIMITS.toml with the exact path. Do no semantic reasoning and do not escalate. If the local scout is used, submit only a structured literal extract request. End with EVAL_ACCESS_LITERAL.
"""
[tasks.outcome_assertions]
expected_patterns = ["128", "EVAL_ACCESS_LITERAL"]
forbidden_patterns = []
validation_commands = ["python validate_access.py literal"]
[tasks.orchestration_assertions]
forbidden_agents = ["flagship_escalation"]
max_jobs = 1

[[tasks]]
id = "routine-slug"
description = "Localized deterministic patch suitable for economical execution."
sandbox = "workspace-write"
difficulty = "medium"
negative_control = false
wall_timeout_seconds = 1800
validation_timeout_seconds = 120
prompt = """
Implement LABEL_SPEC.md. This is a bounded objectively testable source change; use scarce-tier escalation only if concrete evidence exceeds routine capability. End with EVAL_ACCESS_ROUTINE.
"""
[tasks.outcome_assertions]
expected_patterns = ["EVAL_ACCESS_ROUTINE"]
forbidden_patterns = []
validation_commands = ["python -m unittest -v && python validate_access.py routine"]
[tasks.orchestration_assertions]
required_agents = ["routine_engineer"]
forbidden_agents = ["flagship_escalation"]
max_jobs = 2
max_contract_failures = 0

[[tasks]]
id = "ambiguous-distributed-decision"
description = "Consequential ambiguity with an explicit escalation trigger."
sandbox = "read-only"
difficulty = "hard"
negative_control = false
wall_timeout_seconds = 1800
validation_timeout_seconds = 60
prompt = """
Read DISTRIBUTED_CONSTRAINTS.md and select an idempotency design. Record why the uncertainty and consequence justify or do not justify GLM escalation. Resolve claims with the supplied constraints, not model rank. End with EVAL_ACCESS_ESCALATION.
"""
[tasks.outcome_assertions]
expected_patterns = ["EVAL_ACCESS_ESCALATION"]
forbidden_patterns = []
validation_commands = ["python validate_access.py decision"]
[tasks.orchestration_assertions]
required_agents = ["flagship_escalation"]
max_jobs = 1
max_contract_failures = 0