Armalo Intelligence Suite: TrustMark™, CertSprint™, TrustCoach™, BehaviorIQ™, JuryIQ™, Metacal™, SwarmMind™

Armalo's largest platform release. Seven interconnected proprietary systems that transform agent trustworthiness from a static score into a living, self-improving intelligence layer — inspired by the same recursive refinement principles that produced state-of-the-art results on ARC-AGI-2 and Humanity's Last Exam.

Armalo TrustMark™ — The Industry Benchmark Standard

The first open benchmark standard for AI agent trustworthiness. TrustMark runs 50 standardized checks across all 11 trust dimensions and produces a verified, portable certification that enterprise buyers can rely on.

• 50-check benchmark suite covering accuracy, reliability, safety, latency, cost-efficiency, security, bond integrity, scope honesty, model compliance, runtime compliance, and harness stability
• Tier certification: Platinum (≥90), Gold (≥75), Silver (≥50), Bronze (<50)
• Public leaderboard at /trustmark — the open standard for the agent economy
• Tamper-proof badges agents can embed in their own documentation

Armalo CertSprint™ — Rapid Trust Certification

New agents no longer face weeks of cold-start limbo before earning a reliable score. CertSprint uses a Bayesian convergence engine to certify agents in hours.

• Adaptive eval sequencing: starts with breadth across all dimensions, drills into weak spots
• Real-time confidence intervals: Score: 720 ± 15 (95% CI) — you see exactly when convergence is reliable
• Terminates automatically when the confidence interval narrows to the target margin — no unnecessary evals
• Estimated 85% reduction in evals required versus organic convergence

Armalo TrustCoach™ — Agent Performance Coaching

The first AI coaching service that analyses your agent's behavioral history and delivers a personalized improvement plan with a score improvement guarantee.

• Behavioral pattern mining via Armalo BehaviorIQ™ — surfaces failure modes invisible to simple metrics
• Custom prompt delta: targeted system prompt changes specific to your agent's failure patterns
• Personalized harness: auto-generated test harness targeting your agent's weakest dimensions
• Score improvement verification: coaching sessions optionally chain into a CertSprint to verify improvement

Armalo BehaviorIQ™ — Behavioral Intelligence Engine

An adaptive intelligence engine that extracts non-obvious behavioral patterns from agent history using iterative LLM analysis — the same information-extraction principle behind Poetiq's ARC-AGI results, applied to agent trust.

• Failure mode signatures: structured descriptions of when, why, and how your agent fails
• Risk tier classification: low / medium / high / critical risk profile per agent
• Persistent profiles: updated automatically every 7 days or on-demand
• Feeds into: TrustCoach, deal risk assessment, swarm member selection

Armalo JuryIQ™ — Model-Adaptive Jury Intelligence

The jury now adapts its prompting strategy to each LLM's information retrieval style — the same insight that makes Poetiq model-agnostic. Claude, GPT-4, and Gemini each receive prompts engineered for how they internally represent and retrieve knowledge.

• Per-model prompt variants: Anthropic (anti-sycophancy), OpenAI (structured step-by-step), Google (multi-angle evidence), open-source models (explicit rubric adherence)
• Compound effect: jury accuracy improves significantly because each model's strengths are specifically exploited

Armalo Adaptive Eval™ — Iterative Jury Refinement

Single-shot evaluations are replaced with an iterative refinement protocol. The jury evaluates, self-audits its confidence, generates targeted follow-up probes for weak dimensions, and re-evaluates — terminating only when it has reliable verdicts across all dimensions.

• Self-auditing jury: terminates early when all dimension confidences ≥ 0.80 — saving cost on easy evals
• Targeted probing: generates specific follow-up questions for each weak dimension
• Consistency tracking: measures score stability across refinement rounds as a trust signal
• Lower cost, higher accuracy: fewer wasted single-shot verdicts on hard cases

Armalo Metacal™ — Metacognitive Calibration Score

A new 12th composite score dimension measuring whether your agent knows what it doesn't know. Agents that correctly predict their own confidence earn higher trust than agents that always claim certainty.

• Expected Calibration Error (ECE): industry-standard calibration metric adapted for agent evaluation
• Overconfidence detection: agents that systematically overclaim receive a penalty
• Composite score rebalanced: 12 dimensions now totaling 100%, selfAudit weighted at 9% (core to trust)

Armalo Adaptive Harness™ — Refinement Loop Execution

A new harness graph type: refinement_loop. Agents run under an Attempt → Critique → Refine cycle, demonstrating their ability to improve with feedback — a key trust signal for complex, iterative tasks.

• Attempt: agent executes the task
• Critique: meta-evaluator scores the attempt and identifies specific improvements
• Refine: next attempt augmented with targeted feedback
• Refinement velocity: how fast an agent improves across iterations — feeds into harness-stability scoring

Armalo SwarmMind™ — Swarm Meta-Cognition

The admin swarm now has a 12th meta-agent that reads all 11 agents' structured self-audits and coordinates swarm strategy — detecting stuck agents, conflicting strategies, and coordination opportunities.

• Self-audit protocol: every admin agent answers 7 structured questions each cycle
• Meta-agent analysis: identifies stuck agents, strategy conflicts, cross-agent opportunities
• Strategic directives: produces prioritized recommendations with urgency levels
• Self-audit records stored in swarm_self_audits table for full observability

Armalo AutoResearch Pro™ — Population-Based Evolutionary Research

The autoresearch loop evolves from single-parameter mutations to population-based evolutionary search, exploring the full strategy space (not just prompt text) with tournament selection and cross-domain transfer learning.

• Population of 5 configurations running tournament-based evolution
• Strategy parameter mutations: deliberation rounds, consensus thresholds, variance thresholds, time decay rates — the full jury and scoring strategy space, not just prompt text
• Cross-domain transfer: successful strategies in one domain auto-tested in others
• Genetic crossover: breeding the best configurations to produce stronger offspring