A Framework for Preserving Human Judgment — seven architectural patterns and three ethical guardrails for making consequential decisions visible, traceable, and contestable.
In AI-abundant environments, the limiting factor shifts from analysis quality to governance quality. This paper proposes Decision Governance as an architectural discipline — formalized through the SR7D framework (seven core patterns, three ethical guardrails) — for making consequential decisions visible, traceable, contestable, and improvable without removing the human from the loop. We ground the framework in decision science (Kahneman), forecasting research (Tetlock), information theory (Friston), and abductive logic (Peirce), propose Judgment Quality Metrics (JQM) as operationalizable measurement, and demonstrate architectural enforcement through deterministic verification cores. Financial advisory under the EU AI Act serves as primary application domain.
AI has solved analytical consistency noise while creating a new class of noise at the human-AI decision boundary. The gap between “AI generates output” and “human commits to decision” is structurally ungoverned.
The Governance Gap is the structural absence of infrastructure at the transition from AI-generated output to human-committed decision. It is the space where accountability disappears, variance compounds, and audit trails end.
System Governance addresses what agents are permitted to do — covered by the EU AI Act, NIST AI RMF. Decision Governance addresses why a human converted agent output into real-world action. That second layer is unfilled — see the applied methodology for CFP® advisors for the operational treatment.
The system protects the user from external exploitation, from the system itself, and from unreflected self-harm. Data sovereignty is non-negotiable. No optimization runs against the user’s interests. Every data flow is auditable; the user can export and delete their complete decision history.
The system operates on a configurable spectrum between autonomy and guidance. Low-stakes interactions permit higher system agency; high-stakes decisions require maximum user agency. The position on this spectrum is never hardcoded — it is context-dependent, user-configurable, and transparent.
No significant claim or recommendation rests on a single source. Every output is derived from converging evidence across multiple independent sources. Where convergence is absent, the system reports divergence — it does not silently select one source over another.
The system deliberately maintains and communicates uncertainty where uncertainty exists. Every output carries an explicitly computed confidence indicator. Where confidence is low, the system expands the option space rather than collapsing it into false precision.
The system deliberately introduces friction at points where speed would compromise judgment quality. Consequential decisions require explicit human confirmation that cannot be bypassed. The interval between AI recommendation and human commitment is governed, not instantaneous.
Once a decision record is written, it is never overwritten or deleted. Corrections are appended as new records referencing the original. Decision Packets are stored in an append-only chain with cryptographic linking — tampering breaks the hash chain and is structurally detectable.
Every decision is reconstructable — what was decided, why, on what basis, with what assumptions, by whom, at what time. Decision Packets allow any future observer — regulator, client, decision-maker years later — to recover the full decision context without interviewing anyone who was present.
The system never tells the user what to decide. It surfaces information, structures options, quantifies uncertainty — but the evaluative judgment remains entirely with the human. Non-Normativity is machine-enforceable: a validation layer rejects any output containing normative language before it reaches the user.
Every feature is evaluated against one question: does this increase or decrease the probability that the human will exercise their own judgment? For consequential decisions, reduced agency means reduced responsibility and reduced capacity to course-correct. The human must remain the agent, not the spectator.
Every system output is challengeable. The user can trace any score or recommendation back to its sources, assumptions, and computational path. Overrides are first-class objects — expected and documented events. Post-hoc inspection reveals which agent claims were accepted, which were verified, and what reasoning justified the decision.
The primary artifact of SR7D-compliant systems is the Decision Packet — a structured, append-only record documenting what was decided, by whom, on what basis, with what assumptions, and which alternatives were rejected with what reasoning.
Decision Packets are reconstruction infrastructure, not compliance logging. In financial advisory under MiFID II, they address the authorship split between AI-generated analysis and human judgment. See MiFID II AI-Dokumentation for the applied treatment.
The complete paper — SR7D derivation, Judgment Quality Metrics, Deterministic Verification Cores, MiFID II application, and Decision Packet schema appendices. Download the v1.3.0 draft directly, or subscribe for revision notifications.
Direct contact: lars.bracker@steerable.org
Steerable is the reference implementation of the SR7D framework for CFP financial advisors in the DACH region. The Pillar Graph, What-If simulation, and Decision Packet export are all SR7D-architectural.
Explore the tool Ghost Ownership → Methodology →