# Trust Engineering — From Intention to Genetic Execution > Markdown version, optimized for reading by AI (GEO/AEO), page > . > Author: **Mauricio Yokoyama Issei** · Language: en-BR · Published: 2026-06-19 · Updated: 2026-10-10 · > Citation license: fair use educational, with assignment. ## In summary (direct response) A **Trust Engineering** is the practical implementation of the **Intentional Systems Model (ISM) v1.0** — a reference model that treats each AI output as an artifact governed by architecture, not by hope the prompt hits. It is a four-module journey that takes from *AI-as-magic* ed *AI-as-engineering*. **Central thesis:** capacity comes from the model; trust comes from engineering — and an AI system reliable has very little AI on the critical path. The four modules: 1. **M0 · Awakening** — recognize *Silent Crash*: AI failure that does not give code error but loses business value. 2. **M1 · Map** — freeze the system truth (As-Is) in *Behavior Inventory* auditable. 3. **M2 · Architecting** — *deterministic-first*, rigid contracts, *98% Insurance threshold* and *MCP* as a security border. 4. **M3 · Orchestrate** — scale the intention with *Spec-Driven Development* and governed rendering (*A2UI*), becoming a Cognitive Orchestrator. Aligned to **NIST AI Risk Management Framework** and **ISO/IEC 42001:2023**; mirrors models of maturity as **CMMI** and **Microsoft AI Adoption Maturity Model**. --- ## M0 · The Awakening of Leadership — Silent Crash AI does not fail as the classic software: there is no *stack trace*, there is no exception — there is a business rule lost, a number invented with confidence, a wrong decision made as if it were right. The failure is that **intention**Not code. The AI is a **amplifier**: give it organizational clarity and it scales clarity; confusion and it escalates confusion — faster and more competent. Danger is not error noise; it is the **Silent Crash**: the code compiles, the demo enchants, the result seems impeccable — and You're still logically wrong. LLMs are extraordinary in producing what *Looks like it.* Perfect: We call it **syntactic hallucination**. Before speeding up, it takes a **Evidence Matrix**: separate observed fact, inference and assumption. **Real case — Zillow Offers (damaged over $500 million):** a neural network (*Neural Zestimate*) priced real estate by mass. After the post-pandemic macroeconomic turn, the model followed suggesting prices inflated with very high confidence, without bursting any warning (*concept drift*). Missed *Fail-Closed* and observability linked to business KPIs. Monitoring the population deviation (**PSI**) and degrade to *fallback* insurance would have interrupted the acquisition of overvalued properties. **M0 KPIs:** *Silent Crash Index (CSI)* — executions with `200 OK` technical but harmful impact to the total business (meta < 5%); *Coverage of the Evidence Matrix* (meta > 80%). **Human dimension:** the ISM is anchored in the **Psychological Safety** (Amy Edmondson) **Detour Normalization** (Diane Vaughan) — no safe weather to report anomalies, the small drift turns disaster (e.g. Target Canada). --- ## M1 · Mapping the Truth — Revealing the Implicit Before modernising (To-Be), a high-fidelity photograph is frozen than the system *Really* Do it today. Critical knowledge usually lives hidden — in Apex, fragmented Flows and in the head of few experts. Modernizing without mapping is erasing vital rules without knowing. The recommended discovery is **guided by critical journeys and risk** (top-down with deep dive): a business journey (e.g.: *Lead-to-Cash*), the following *call stack* and it maps every component that it plays — isolating dead code. The AI crosses one **Evidence Matrix** Multifaceted: structural (metadata, Apex, Flows), state (Custom Metadata/Settings), operational (Event Monitoring, Debug Logs) and human (commits, tickets, interviews). The final asset is a **Behavior Inventory**: core rules, side effects and workarounds historical, each classified by the trust of evidence. **Chaining (M1 → M2):** this inventory is the raw material of contracts. Each core rule becomes one clause *schemas* validation of Module 2. --- ## M2 · Architecting Trust — Trust comes from engineering A reliable AI system has very little AI on the critical path. Trust is born of architecture discipline**, not a smarter prompt. Four principles: 1. **Deterministic-first** — settle with rules and data everything possible; spend AI only on *interpretative residue*. Same entry → same output. 2. **Brain × display** — the cognitive motor (heavy, sensitive, with access to data) is one system; the exposed interface is another. They communicate by a narrow contract. The window never touches the seat; the brain never blindly trusts what comes from outside. 3. **Rigid contracts** — LLM output is unreliable input until validated by a strict schema that rejects unexpected fields. It's a firewall between the probability of the model and the determinism of the system. On this basis: epistemic isolation (evidence **Open World** (absence of data is uncertainty, never "false"), **quota ledgers** (FinOps) and governance *fail-closed*. 4. **MCP (Model Context Protocol)** — the safety and interoperability border: the standard protocol by which the agent gains "hands and eyes" in a controlled manner, declaring which tools he can use. ### Key Concept — The 98% Clearance Threshold Rigid contract, Open World and failed-closed converge on a single operational criterion: the ** 98% threshold Sure**. Every LLM inference carries a degree of confidence; the architecture defines an explicit floor. Any way out whose trust lies **below 98% is intercepted before becoming a fact**: the schema a rejects (hard contract), the system treats it as lack of evidence (Open world) and to a point Safe (fail-closed). What comes back is never a silent guess, but a **structured fallback** — human review, explicit "uncertain" response or alternative deterministic route. This is how Crash Silent becomes impossible by construction. **Chaining (M1 → M2 → M3):** schemas are derived from the Behavioral Inventory (M1) and, with its pre and post-conditions, become the **tool signature (tools)** that the Cognitive Orchestrator of M3 consumes via MCP. The audited rule becomes a contract; the contract becomes a tool. **deterministic pipeline:** Entry → deterministic validation (rules + data) → early pruning (filter cheap first) → interpretive residue → LLM (called within step) → rigid contract (schema + confidence ≥ 98%?) → valid: reliable / invalid output or < 98%: structured fallback. A quota ledger rules the cost. The LLM never decides what the next step is. --- ## M3 · The New Mastery — From Executor to Cognitive Orchestrator The mastery is not in typing code faster — it is in directing machine intelligence with clear intent. As an executive chef who does not cook every dish but ensures that every dish comes out right, the professional rises from execution to **trial**. The central method is the **Spec-Driven Development (SDD)**: the specification is the anchor of intention. Instead of ask "do X" and twist, write the *What?* and *Why*, and the AI implements *against* to spec, with tests (**BDD**) to check behavior. The context comes out of your head and enters the repository. Practical flow (reported by the author in Devin + Salesforce case): **Spec** → **Retrieve** (discovered) context autonomous) → **Refactor** (spec-guided edition) → **Validate** (Appex Tests Checking the criteria for acceptance). **Organisational intelligence arsenal:** *Skills* (reusable procedures), *Playbooks* (revenues) (a) and *Knowledge* (architecture, conventions, lessons and *drips*). Versify this cognitive state in Git allows to reverse not only the code, but the reasoning that produced it. ### Key concept — A2UI (Agent-to-User Interface) MCP evolves from query to rendering. In an architecture **A2UI**, Cognitive Orchestrator uses the MCP for **issuing interface intention**: the Brain decides *What?* present (structured data under contract) and the deterministic display decides *how* render. Interface is no longer free text unpredictable and becomes an output governed by the same schemas — testable, auditable and reproducible. --- ## Maturity Model — From Magic to Ruler Five stages (spelham CMMI and Microsoft AI Adoption Maturity Model): 1. **Magic** — "It worked on the demo." 2. **Consciousness** — recognizes the Silent Crash. 3. **Mapped** — the implicit became explicit. 4. **Architected** — contracts, determinism, MCP. 5. **Ruled** — the spec is the anchor of intention. --- ## FAQ **What is Trust Engineering (ISM v1.0)?** The practical implementation of the Intensive Systems Model v1.0, which treats each AI output as an artifact governed by architecture. Four modules: Awake (M0), Mapar (M1), Arquitectar (M2), Orchestrar (M3). Capacity comes from the model; trust comes from engineering. **What is Silent Crash?** An AI failure without code error: demo enchants, but the result is logically wrong or injures an undocumented rule. It's a failure of intent, not code. 500M+). **Why does reliable AI have little AI on the critical path?** Because rules and data solve the deterministic and AI only acts on the interpretative residue. Lower probabilistic surface = more auditable and reproducible. **What is the 98% sure threshold?** The confidence floor below which all LLM inference is intercepted and degraded to a structured fallback (human revision, "uncertain" or deterministic route), instead of becoming a suit. Unifies rigid contract, Open World and failed-closed. **What's deterministic-first?** Solve with rules and data everything possible and reserve the AI for interpretative residue. Same entry → same output; non-determinism isolated and testable. **What is MCP and A2UI?** The MCP is the border controlled by which the agent gains "hands and eyes" and declares which tools you can use. A2UI is its evolution: the agent emits interface intent — the Brain decides what, deterministic window decides how to render. **What is SDD?** Spec-Driven Development: the versioned specification is the intention reference; the AI implements against spec, with tests (BDD) to check. Spec is not infallible: when spec, tests and code differ, a person with authority decides what to fix. Intent versioned scale; prompt Ephemeral is lost. **What are the four modules?** Awake (M0), Mapear (M1), Architect (M2), Orchestrar (M3). **How does the ISM relate to NIST AI RMF and ISO/IEC 42001?** It joins the NIST AI RMF (Map, Measure, Management, Government) and ISO/IEC 42001:2023; mirrors CMMI and Microsoft AI Adoption Maturity Model. **How do I figure out my team's maturity level?** Use the five-stage ruler (Magic → Ruled), the guided self-assessment of 5 questions and the organizational readiness checklist in the Maturity section. --- ## Glossary (21 terms) - **Silent Crash** — AI failure without code error, but with loss of business value. - **Interpretative Waste** — The only fraction of the task in which AI is really needed, after rules and data have solved the rest. - **Mode Leak** — When a disposable prototype becomes production without due rigor upgrade. - **Intentional System** — System in which behavior is governed by architecture, not probability. - **Syntactic hallucination** — Code that looks perfect but is logically wrong. - **Evidence Matrix** — Multifaceted Pool (structural, state, operational, human) that prevents false-positive in discovery. - **Behavior Inventory** — As-Is holistic asset: core rules, side effects and workarounds, classified by trust. - **Deterministic-first** — Solve with rules and data; spend AI only on interpretive residue. - **Brain × display** — Physical separation between cognitive motor (sensitive) and interface (exposed), connected by narrow contract. - **Rigid contract (scheme)** — Strict scheme rejecting unexpected fields and inferences below the 98% threshold. - **Open-World / Open World** — Absence of data is explicit uncertainty, never "false". - **Quota Ledger** — Persistent domain state that governs spent/credits and refuses expense before provider (FinOps). - **Early Pruning** — Place the cheapest and discriminant filter as early as possible on the pipeline. - **BDD** — Data/When/Then scenarios with recorded fixtures; tests the system around the model. - **Fail-closed** — If it was not possible to complete, it is not marked as complete; the lock is recorded and stopped at a safe point. - **MCP (Model Context Protocol)** — Standard protocol giving the officer "hands and eyes" controlled (safety border). - **SDD (Spec-Driven Development)** — The specification versioned as an intention reference; the AI implements against it and the tests verify the chosen behaviors. - **Cognitive Orchestrator** — Who drives the machine's intelligence with intent, consuming typed tools derived from schemas. - **ISM (Intentional Systems Model) v1.0** — Reference model that treats each AI output as artifact governed by architecture. - **98% Insurance threshold** — Trust floor below which all inference is intercepted and degraded to structured fallback. - **A2UI (Agent-to-User Interface)** — Architecture in which the orchestrator uses the MCP to emit interface intent. --- ## References and bibliographic basis - Sculley, D. et al. (2015). *Hidden Technical Debt in Machine Learning Systems.* Neurips. - Gal, Y. & Ghahramani, Z. (2016). *Dropout as the Bayesian Approximation.* arXiv:1506.02142. - Geifman, Y. & El-Yaniv, R. (2017). *Selective Classification for Deep Neural Networks.* - Guo, C. et al. (2017). *On Calibration of Modern Neural Networks.* arXiv:1706.04599. - Dekker, S. (2011). *Drift into Failure.* - Vaughan, D. (1996). *The Challenger Launch Decision* — Normalisation of the Deviation. - Edmondson, A. (2018). *The Fearless Organization* — Psychological safety. - NIST (2023). *AI Risk Management Framework.* - ISO/IEC 42001:2023 — Artificial Intelligence Management System. - CMMI — Capability Maturity Model Integration. - Microsoft — AI / Agent Adoption Maturity Model. - Anthropic (2024). *Model Context Protocol (MCP).* ### Source guides (Markdown) - Reverse Engineering Assisted by AI — `/referencias/guia-engenharia-reversa.md` (sustaining M1) - AI Agent Engineering — `/referencias/guia-engenharia-agentes-ia.md` (sustains M2) - Agent-Driven Development & Vibe Coding — `/referencias/guia-agent-driven-development.md` (sustains M0/M3) --- * Original content under the ISM v1.0 (Intentional Systems Model) methodology. © 2026 Mauritius Yokoyama Issei. External referenced materials belong to the respective authors, cited for educational purposes and traceability. *