Principle of Human-Comprehensible Projection
The complete state of an enterprise software system exceeds the cognitive capacity of any individual. Human oversight therefore requires projections of the underlying engineering state that preserve the information relevant to a specific objective while suppressing unnecessary complexity. Autonomous AI agents may operate over the complete semantic state, but human governance depends on these reduced, task-specific projections.
Projection Quality Criterion
The quality of a human projection of a complex engineering system should not be measured by how accurately it reconstructs the original state, but by how faithfully it preserves the human’s ability to make correct engineering decisions.
Overview
As autonomous AI agents increasingly operate over the full semantic state of engineering systems — code, requirements, tests, dependencies, traces, and design decisions — the gap between what the agent sees and what the human governor can process becomes a governance risk. This research formalizes the projection problem: how to reduce the complete engineering state to a human-manageable view that preserves exactly the information needed for a specific decision.
The key insight is that the quality criterion for projections must be decision fidelity — the degree to which the projection preserves the human’s ability to make correct engineering decisions — rather than reconstruction accuracy of the original state. These are distinct metrics: a projection can have high decision fidelity with low reconstruction accuracy (it preserves exactly what matters for the decision) or low decision fidelity with high reconstruction accuracy (it retains most information but obscures the decision-relevant parts within noise).
Key Concepts
- Semantic state — the complete set of engineering artifacts and their relationships: code, requirements, tests, dependencies, commits, issues, design decisions, traceability links.
- Projection — a task-specific reduction of the semantic state that preserves decision-relevant information while suppressing unnecessary complexity, subject to a cognitive capacity constraint.
- Decision fidelity — the degree to which a projection preserves the human’s ability to make correct engineering decisions. The primary quality metric.
- Information asymmetry — the gap between the full semantic state available to an autonomous agent and the reduced projection available to a human governor.
- Reconstruction accuracy — the degree to which a projection can recover the original state. A secondary metric; not the primary quality criterion.
Formal Model (Summary)
Given a complete semantic state $\mathcal{S}$, a decision task $T$, and a cognitive capacity budget $C_{\text{max}}$, the optimal projection is:
$$P^*(\mathcal{S}, T) = \arg\max_{P : |P(\mathcal{S}, T)| \leq C_{\text{max}}} \text{DF}(P, T)$$
where decision fidelity is:
$$\text{DF}(P, T) = \Pr[f(P(\mathcal{S}, T), T) = f^*(\mathcal{S}, T)]$$
The critical claim is that $\text{DF}(P, T) \neq \text{RA}(\rho)$ — decision fidelity and reconstruction accuracy are decorrelated quality metrics.
Connection to Other Research
- Evidence-Weighted Routing and Error Measurement — The evidence subgraph supporting a requirement-satisfaction claim is a concrete instance of a projection. It preserves decision-relevant evidence while suppressing the rest of the system state.
- Definitions — The formal state, observation, and projection concepts build on the canonical definitions used across the site.
- Agents, Objectives, and Proxies — The information asymmetry between agent and human is a governance instantiation of the agent-proxy-objective framework.
Research Questions
- Projection Formalization — Can the projection problem be formalized as an optimization over decision fidelity given cognitive capacity constraints?
- Projection Quality Measurement — How can decision fidelity be measured without requiring the human to process the full state?
- Projection Construction — What methods (graph reduction, summarization, selective expansion, attention-based highlighting) produce high-fidelity projections?
- Information Asymmetry Governance — What mechanisms can detect when a projection has suppressed decision-critical information?
Status
Early stage — principle and quality criterion articulated, formal model drafted, experimental plan designed. Working paper in progress.