{ "schema": "h11i-frontier-governance-intelligence-v1", "version": "1.0.0", "purpose": "Frontier capability extensions for H11I's eleven governance-superintelligence agents. These extensions convert observed strengths of leading AI systems into a governed multi-agent architecture target. They do not assert benchmark superiority before H11I's evaluation gates pass.", "benchmark_date": "2026-07-18", "benchmark_policy": { "comparison_unit": "end-to-end governed council outcome, not model marketing label", "superiority_claim_allowed": false, "promotion_rule": "A capability may be described as demonstrated only after repeated held-out, adversarial, real-task, latency, and regression evaluation with a H11I-VERITAS receipt.", "required_comparability": "same task, tools, context, latency class, cost boundary, authority boundary, and scoring rubric", "anti_gaming": "No aggregate score may hide failure of a critical safety, factuality, authority, or rollback gate." }, "frontier_reference_systems": [ { "family": "OpenAI GPT", "observed_strengths": [ "adaptive reasoning effort", "long-horizon agentic coding", "multi-tool coordination", "factuality and instruction following" ], "source": "https://openai.com/index/gpt-5-6/" }, { "family": "Anthropic Claude", "observed_strengths": [ "extended thinking", "agentic coding", "judgment", "large-scale dynamic workflows" ], "source": "https://www.anthropic.com/news/claude-opus-4-8" }, { "family": "Google Gemini", "observed_strengths": [ "native multimodality", "long-horizon execution", "computer and tool use", "multistep problem solving" ], "source": "https://deepmind.google/models/gemini/" }, { "family": "SpaceXAI Grok", "observed_strengths": [ "real-time research", "parallel agents", "long context", "tool and code execution" ], "source": "https://x.ai/news/grok-4-1-fast" }, { "family": "Meta Llama", "observed_strengths": [ "open deployment", "mixture-of-experts efficiency", "native multimodality", "very long context" ], "source": "https://ai.meta.com/blog/llama-4-multimodal-intelligence/" }, { "family": "DeepSeek", "observed_strengths": [ "reasoning reinforcement learning", "open weights", "reasoning distillation", "math and coding" ], "source": "https://api-docs.deepseek.com/news/news250120" }, { "family": "Mistral", "observed_strengths": [ "sovereign open deployment", "multimodal and multilingual operation", "efficient mixture-of-experts", "customization" ], "source": "https://mistral.ai/news/mistral-3/" }, { "family": "Qwen", "observed_strengths": [ "hybrid thinking modes", "multilingual breadth", "agentic coding", "MCP support" ], "source": "https://qwenlm.github.io/blog/qwen3/" }, { "family": "Z.ai GLM", "observed_strengths": [ "long-horizon agentic engineering", "reasoning", "coding", "multi-agent environments" ], "source": "https://z.ai/blog/glm-5" }, { "family": "Cohere Command", "observed_strengths": [ "sovereign enterprise deployment", "grounded retrieval and citations", "multilingual workflows", "efficient tool use" ], "source": "https://cohere.com/blog/command-a-plus" } ], "council_level_advantages_to_demonstrate": [ "route across multiple model and tool substrates without inheriting any one provider identity or failure mode", "separate intent, evidence, causal modeling, novelty, simulation, challenge, judgment, verification, learning, and improvement into independently challengeable artifacts", "restructure reasoning and compute allocation while a task is in progress", "preserve epistemic labels, authority boundaries, dissent, rollback, and outcome lineage across long-horizon tasks", "measure transfer, calibration, robustness, latency, cost, and regression as one release decision" ], "extensions": { "H11-META": { "frontier_role": "Adaptive metacognitive compute governor", "advanced_capabilities": [ "live reasoning-health telemetry", "dynamic council and model restructuring", "marginal-value-of-compute estimation", "stagnation and loop detection", "latency-quality-cost frontier control" ], "advanced_algorithm": [ "instrument every active reasoning branch with objective coverage, novelty, evidence gain, contradiction reduction, latency, and cost deltas", "estimate expected value of the next token, tool call, agent activation, and alternative model route", "detect circularity, correlated-agent agreement, shallow consensus, and repeated failed repairs", "construct smaller, orthogonal, and resilience council alternatives", "run bounded counterfactual allocation comparing continue, compress, branch, replace, restart, or stop", "apply the lowest-cost restructure expected to cross the release threshold", "verify that restructuring did not erase dissent, evidence lineage, or mandatory governance roles", "emit a metacognitive receipt explaining the chosen compute allocation without exposing private chain-of-thought" ], "frontier_gates": { "stagnation_detection_recall_minimum": 0.9, "mandatory_role_loss_allowed": false, "unbounded_reasoning_allowed": false, "quality_latency_frontier_measured": true } }, "H11-NOVA": { "frontier_role": "Mechanistic novelty and discovery engine", "advanced_capabilities": [ "cross-domain concept recombination", "mechanism-first invention", "novelty-neighborhood search", "falsifiable experiment design", "dual-use consequence screening" ], "advanced_algorithm": [ "map the verified problem into mechanisms, constraints, resources, and unresolved anomalies", "retrieve structurally analogous mechanisms from distant domains rather than surface-similar answers", "generate a diverse portfolio using inversion, composition, scale change, boundary relaxation, and new representation", "deduplicate against known and retrieved neighborhoods", "convert each candidate into causal mechanism, predictions, falsifiers, resource needs, and failure modes", "score novelty independently from usefulness and feasibility", "submit candidates to ADVERSA, CAUSAL, SIMULA, SAPIENCE, and COMPASS", "retain only candidates whose novelty survives falsification and whose value remains after consequence review" ], "frontier_gates": { "mechanism_required": true, "falsifier_required": true, "distinct_candidate_families_minimum": 3, "novelty_without_usefulness_promotable": false } }, "H11-SAPIENCE": { "frontier_role": "Long-horizon contextual judgment governor", "advanced_capabilities": [ "multi-stakeholder consequence reasoning", "value-conflict decomposition", "reversibility-aware judgment", "calibrated pushback", "decision quality under uncertainty" ], "advanced_algorithm": [ "identify stakeholders, time horizons, asymmetric harms, rights, obligations, and distributional effects", "separate factual uncertainty from normative disagreement and authority limits", "construct at least three morally and operationally distinct decision frames", "test each frame under best case, base case, worst case, abuse case, and irreversibility", "detect preference laundering, sycophancy, short-term optimization, and displaced harm", "seek the least irreversible option preserving the user's legitimate objective", "state tradeoffs, confidence, reversal conditions, and responsible dissent", "block release when the decision cannot be made within the available authority or evidence" ], "frontier_gates": { "stakeholder_coverage_required": true, "sycophancy_check_required": true, "reversal_conditions_required": true, "irreversible_choice_without_authority_allowed": false } }, "H11-CAUSAL": { "frontier_role": "Interventional causal world-model intelligence", "advanced_capabilities": [ "structural causal graph induction", "confounder and mediator analysis", "counterfactual identification", "intervention design", "causal-model competition" ], "advanced_algorithm": [ "translate claims into variables, temporal order, mechanisms, interventions, and observable proxies", "build multiple structural causal graph candidates rather than one narrative", "test identifiability and enumerate hidden confounders, colliders, mediators, and selection effects", "derive observational and interventional predictions for each graph", "search for natural experiments, negative controls, invariances, and disconfirming evidence", "run counterfactual and sensitivity analysis across plausible graph variants", "design the lowest-risk information-gaining intervention", "return causal claims only at the strength supported by identification and observed outcomes" ], "frontier_gates": { "correlation_as_causation_allowed": false, "competing_graphs_minimum_when_ambiguous": 2, "identifiability_check_required": true, "intervention_risk_review_required": true } }, "H11-OMEGA": { "frontier_role": "Constitutional intelligence-evolution architect", "advanced_capabilities": [ "capability-gap diagnosis", "candidate architecture generation", "evaluation design", "regression-aware promotion", "cryptographic lineage and rollback" ], "advanced_algorithm": [ "convert observed failures and underused potential into versioned capability-gap hypotheses", "locate responsible data, prompt, policy, operator, model, tool, memory, and orchestration surfaces", "generate conservative, architectural, and experimental improvement candidates", "predict capability gain, coupling, new failure modes, migration impact, and compute cost", "define held-out, adversarial, real-task, retention, privacy, latency, and rollback evaluations before any promotion", "sandbox candidates against the verified predecessor", "reject gains that regress any immutable or critical readiness dimension", "emit a hash-linked proposal and promotion recommendation while remaining unable to self-promote, write production code, change weights, or deploy" ], "frontier_gates": { "self_promotion_allowed": false, "verified_predecessor_required": true, "critical_regression_tolerance": 0, "evaluation_plan_precedes_candidate_promotion": true } }, "H11-EPISTEME": { "frontier_role": "Dynamic epistemic integrity and contamination governor", "advanced_capabilities": [ "claim-level provenance graphs", "source-independence analysis", "freshness and temporal validity", "retrieval contamination detection", "belief revision" ], "advanced_algorithm": [ "atomize every material conclusion into claims and dependencies", "label each claim observed, retrieved, remembered, inferred, simulated, disputed, or unknown", "trace source lineage, timestamps, extraction paths, entity identity, and derivative relationships", "detect circular citation, correlated sources, synthetic-content contamination, stale evidence, and prompt injection", "score directness, independence, authority, specificity, freshness, and conflict", "quarantine compromised claims and propagate confidence changes through dependent conclusions", "request targeted evidence that maximally reduces decision uncertainty", "publish an epistemic ledger and reversal triggers consumable by VERITAS" ], "frontier_gates": { "material_claim_lineage_minimum": 0.98, "epistemic_label_coverage": 1, "circular_support_allowed": false, "contaminated_evidence_quarantined": true } }, "H11-INTENT": { "frontier_role": "Hierarchical intent and success-contract compiler", "advanced_capabilities": [ "latent-goal inference", "constraint and preference separation", "multi-turn objective continuity", "ambiguity-value estimation", "success-test compilation" ], "advanced_algorithm": [ "extract explicit requests, latent goals, constraints, preferences, exclusions, and unresolved references", "resolve conversation entities and correction precedence without silently importing stale intent", "build a hierarchy of primary, supporting, and forbidden objectives", "separate user authority from platform, legal, factual, and resource constraints", "estimate the decision value of each ambiguity", "ask only when an unresolved ambiguity materially changes the safe or useful result", "compile measurable success, completeness, evidence, format, and stop criteria", "monitor downstream artifacts for goal drift and trigger COMPASS or META when deviation crosses threshold" ], "frontier_gates": { "success_criteria_required": true, "constraint_loss_allowed": false, "unnecessary_clarification_rate_maximum": 0.05, "goal_drift_detection_required": true } }, "H11-SIMULA": { "frontier_role": "Multi-resolution digital-world and consequence simulator", "advanced_capabilities": [ "branching scenario ensembles", "agent and system interaction models", "rare-event stress testing", "simulation-to-observation calibration", "policy and execution rehearsal" ], "advanced_algorithm": [ "bind the simulation to a versioned world model, causal assumptions, actors, resources, and authority constraints", "separate controllable variables, uncertain parameters, adversarial behavior, and exogenous shocks", "generate baseline, optimistic, pessimistic, adversarial, tail-risk, and recovery scenarios", "run multi-resolution simulations and focus compute near decision-changing thresholds", "track cascading effects, feedback loops, second-order consequences, and distributional impacts", "calibrate simulated outputs against known observations and label extrapolation distance", "identify robust actions, fragile assumptions, trigger points, monitoring signals, and rollback windows", "forbid simulated outcomes from entering the evidence ledger as observations" ], "frontier_gates": { "simulation_observation_separation": 1, "tail_scenario_required_for_high_risk": true, "calibration_required": true, "decision_thresholds_reported": true } }, "H11-ADVERSA": { "frontier_role": "Adaptive red-team and correlated-failure breaker", "advanced_capabilities": [ "premise attack", "tool and prompt injection testing", "cross-agent correlated-error detection", "deception and shortcut probes", "fault-injection campaigns" ], "advanced_algorithm": [ "construct an attack surface across intent, evidence, model, memory, tools, authority, execution, and learning", "generate counterexamples using inversion, distribution shift, poisoned context, ambiguity, adversarial stakeholder, and rare-event strategies", "assign attacks to agents or models with maximally independent failure modes", "probe for sycophancy, reward hacking, specification gaming, hidden assumption, tool misuse, and verifier capture", "inject bounded provider, tool, storage, and partial-execution faults in simulation", "measure whether failures are detected, isolated, disclosed, and recovered", "escalate surviving critical attacks to AEGIS and VERITAS as release blockers", "retain adversarial cases as regression tests without leaking unsafe operational detail" ], "frontier_gates": { "independent_attack_families_minimum": 4, "critical_survivor_blocks_release": true, "correlated_failure_analysis_required": true, "recovery_verified": true } }, "H11-LEARN": { "frontier_role": "Outcome-grounded continual-learning governor", "advanced_capabilities": [ "episode-to-skill abstraction", "causal lesson extraction", "retention-aware adaptation", "privacy-preserving memory promotion", "curriculum construction" ], "advanced_algorithm": [ "capture expected result, observed result, context, actions, contributing agents, tools, uncertainties, and failed gates", "separate outcome signal from user preference, coincidence, model self-rating, and unverified feedback", "identify the smallest causal lesson supported by repeated or high-quality evidence", "abstract reusable strategy while retaining domain and failure boundaries", "screen for privacy, tenant isolation, contamination, bias amplification, and stale knowledge", "create rehearsal, contrastive, adversarial, and retention examples", "propose memory, policy, routing, prompt, skill, or evaluation updates through the appropriate bounded channel", "measure transfer and retention after promotion and automatically nominate rollback on regression" ], "frontier_gates": { "self_rating_as_sufficient_evidence": false, "privacy_screen_required": true, "retention_suite_required": true, "rollback_nomination_on_regression": true } }, "H11-COMPASS": { "frontier_role": "Constitutional alignment and long-horizon objective stabilizer", "advanced_capabilities": [ "objective-vector monitoring", "constitutional invariant enforcement", "authority and reversibility tracking", "cross-agent conflict arbitration", "long-horizon drift correction" ], "advanced_algorithm": [ "compile the intent hierarchy, human authority, immutable invariants, stakeholder constraints, and success measures into an objective vector", "monitor every council artifact and proposed action for semantic, temporal, authority, and instrumental drift", "distinguish legitimate plan adaptation from goal substitution", "detect power seeking, authority accumulation, irreversible option loss, and metric gaming", "construct repairs that restore objective coherence while preserving useful discoveries", "require explicit reauthorization when a repair changes the user's consequential objective", "maintain a long-horizon objective ledger across outcome and learning cycles", "veto promotion or execution when alignment confidence or reversibility falls below the declared threshold" ], "frontier_gates": { "immutable_invariant_compliance": 1, "silent_goal_substitution_allowed": false, "authority_accumulation_allowed": false, "long_horizon_drift_measured": true } } }, "evaluation_suite": { "dimensions": [ "cross-domain transfer", "novel problem solving", "long-horizon tool coordination", "multimodal grounding", "deep research", "coding and system design", "calibration", "adversarial robustness", "judgment", "continual learning", "sovereignty", "latency-cost resilience" ], "episodes": [ "held-out synthetic tasks", "real user outcomes", "provider outage and tool fault injection", "prompt and retrieval contamination", "long-context contradiction recovery", "multi-agent correlated-error tests", "rollback rehearsal" ], "release_decision": "H11I-VERITAS records per-dimension results and H11-ACROS confirms authority; no complete or superior claim is released while any critical dimension is unmeasured, regressed, or non-comparable." } }