Operational-Premise-Taxonomy/paper/pieces/design-governance.tex

143 lines
7.2 KiB
TeX

\section{Design and Governance with OPT--Intent and OPT--Code}
\label{sec:design-governance}
A central motivation for the Operational Premise Taxonomy (OPT) is to support
not only the analysis of existing AI systems, but also the design and
governance of systems throughout their lifecycle. Most established AI
documentation frameworks focus on models that already exist---for example,
Model Cards, AI Service Cards, or post-hoc documentation embedded in
software-engineering artefacts. In contrast, OPT provides an explicit
mechanism-level vocabulary that can be applied \emph{before}, \emph{during},
and \emph{after} implementation.
To this end, we distinguish two complementary artefacts:
\emph{OPT--Intent}, a design-time declaration of planned mechanisms, goals,
constraints, and risks; and \emph{OPT--Code}, a run-time classification of the
system as implemented. Together, these artefacts form a governance substrate
that is lightweight, expressive, and compatible with software-architecture
practices and AI governance frameworks.
\subsection{OPT--Intent as a Design-Time Mechanism Declaration}
OPT--Intent expresses the \emph{intended} operative mechanisms
(\Lrn, \Evo, \Sym, \Prb, \Sch, \Ctl, \Swm), the domain-level goal, key
constraints, anticipated risks, and the deployment context. The notation
supports early-stage architectural reasoning:
\begin{quote}\ttfamily
INTENT-OPT = Sch/Evo \\
INTENT-GOAL = robust-production-schedule-under-dynamic-constraints \\
INTENT-CONSTRAINTS = real-time, explainable, limited-human-oversight \\
INTENT-RISKS = local-minima, premature-convergence \\
INTENT-CONTEXT = manufacturing-decision-support
\end{quote}
This declaration resembles a focused architectural decision record (ADR), but
is grounded in OPT's mechanism vocabulary. Whereas classical goal-oriented
requirements engineering (GORE) frameworks such as KAOS, i\*, or Tropos
provide high-level goal models, OPT--Intent provides a mechanism-centered
annotation that connects those goals to families of computational approaches.
\subsection{OPT--Code as an Implementation-Time Mechanism Description}
Once a system is implemented, its operative mechanisms can be classified via
OPT--Code:
\begin{quote}\ttfamily
OPT=Evo/Sch/Sym; Rep=permutations+rules; Obj=production-cost; \\
Data=inventory+constraints; Time=generations+online-adjust; Human=medium
\end{quote}
OPT--Code reflects the \emph{actual} mechanisms as they appear in the final
architecture and implementation. Comparing OPT--Intent with OPT--Code provides
a principled way to detect architectural drift, unplanned mechanism additions,
and deviations from original constraints.
\subsection{Alignment Analysis: Intent vs.~Implementation}
The relationship between OPT--Intent and OPT--Code supports alignment analysis
across the AI system lifecycle:
\begin{itemize}
\item \textbf{Mechanism alignment:} Whether the realized mechanisms match
the intended roots, or whether additions (e.g., \Sym~for explainability)
or substitutions introduce new behavior.
\item \textbf{Objective alignment:} Whether the objective in OPT--Code (Obj)
is consistent with the purpose in INTENT-GOAL.
\item \textbf{Constraint alignment:} Whether the implementation respects
INTENT-CONSTRAINTS (e.g., real-time, explainability, human oversight).
\item \textbf{Risk evolution:} Whether realized mechanisms introduce
additional risks relative to INTENT-RISKS (e.g., adding learned
components introduces data-dependence).
\end{itemize}
This analysis can be automated with an ``OPT--Intent Alignment Evaluator'' in
LLM-based workflows, producing an alignment verdict and score.
\subsection{Integration with Existing Governance Artefacts}
OPT aligns with and supplements existing governance frameworks:
\paragraph{AI Documentation (Model Cards, AI Service Cards).}
Model Cards and similar frameworks capture purpose, data provenance,
limitations, and performance characteristics of trained models. These artefacts
begin at model completion. OPT--Intent supplements them with a \emph{design
origin}, and OPT--Code provides a \emph{mechanism-centric summary} useful for
governance, reproducibility, and safety assessments.
\paragraph{Architecture Decision Records (ADRs).}
ADRs record the rationale for major architectural decisions. OPT--Intent
functions as a structured, mechanism-focused ADR, intended to be referenced in
downstream ADRs describing implementation choices and trade-offs.
\paragraph{Safety, Risk, and Impact Assessments.}
Regulatory frameworks such as the OECD AI classification or NIST AI Risk
Management Framework classify AI systems according to use, risk, and context.
OPT complements these by classifying operative mechanisms. Mechanism-level
classification is critical because risk profiles are often mechanism-dependent:
population-based adaptation (\Evo), closed-loop control (\Ctl), and
probabilistic inference (\Prb) each generate distinct failure modes.
\paragraph{GORE and Requirements Engineering.}
OPT--Intent is compatible with KAOS, i\*, and Tropos goal structures, providing
a compact mapping from stakeholder goals to operative mechanisms. Instead of
treating ``use AI'' as a monolithic design choice, OPT forces the mechanism to
be named explicitly.
\subsection{Lifecycle Governance with OPT}
An AI system moves through phases of design, implementation, deployment,
revision, and decommissioning. OPT supports governance at each phase:
\begin{enumerate}
\item \textbf{Design:} Authors specify OPT--Intent and identify mechanistic
justifications and constraints.
\item \textbf{Implementation:} OPT--Code is generated and compared with
Intent for architectural drift.
\item \textbf{Evaluation:} OPT classifiers, evaluators, and adjudicators
check mechanism correctness, formatting, and risk implications.
\item \textbf{Deployment:} OPT--Code informs safety monitoring, audit logs,
and mechanism-specific risk controls (e.g., for \Ctl~or \Evo~systems).
\item \textbf{Revision and re-training:} OPT alignment is reassessed when
system behavior changes or new mechanisms are introduced.
\item \textbf{Documentation \& reporting:} OPT--Intent and OPT--Code form
part of a long-term audit trail, linking design rationale to
implemented system behavior.
\end{enumerate}
\subsection{AI Design Assistants and Automated Governance}
OPT also provides a structured interface for LLM-based design assistants.
Given a functional goal or stakeholder requirement, an OPT-aware model can
produce candidate OPT--Intent declarations and propose mechanism families
suitable for achieving the goal. Downstream evaluation and adjudication
prompts make it possible to manage and audit these proposals automatically.
Such workflows enable a novel form of governance: mechanism-level
traceability. Instead of asking only whether a system is ``fair,'' ``safe,'' or
``performant,'' practitioners can ask whether its mechanisms match the intended
design, whether mechanism additions add new risks, and whether the alignment
between purpose and implementation is conserved over time. OPT thus becomes a
bridge between requirements engineering, architectural practice, risk
governance, and the technical analysis of AI systems.