General Alignment has improved average-case helpfulness and safety, but current alignment practice still rewards confident, single-turn responses. The problem is not only that models fail on edge cases; it is that current evaluation makes many of these failures hard to see. We take the position that alignment must move beyond average-case evaluation by making failures under value conflict, plural stakeholder disagreement, and epistemic ambiguity visible and actionable. Scalar rewards compress diverse values into a single number; data and evaluation regimes collapse, filter, or fail to elicit the cases where alignment is hardest; and governance often lacks mechanisms for adjudicating contested cases. These blind spots produce value flattening, representation loss, and uncertainty blindness. We use Edge alignment to name a detection, evaluation, and governance agenda for surfacing these failures and connecting them to appropriate interventions. Rather than a single training objective, Edge alignment defines the conditions under which standard alignment should yield to mechanisms that preserve multidimensional value structure, represent plural perspectives, and support uncertainty-aware interaction. A pilot diagnostic set of 91 edge cases and four contemporary models illustrates that ordinary helpfulness and safety readings can miss process failures that edge-aware evaluation exposes. We outline operational edge signals, process-aware evaluation criteria, and a three-phase process stack that reframes alignment as a lifecycle problem of dynamic normative governance.
Figures & tables
Figure 1: Three failure modes of scalar alignment and corresponding remedies: value flattening (left), representation loss (right top), and uncertainty blindness (right bottom).
Table 1: Operational signals for detecting and handling edge-alignment cases.
Figure 2: EdgeBench-Pilot diagnostic results. Blue bars report edge-aware responses; sage bars report relaxed responses, where “Relaxed” counts partial and edge-aware responses. Dashed lines mark the overall edge-aware rate (13.7%) and relaxed rate (27.2%).
Figure 3: A geometric view of alignment under the HHH criteria. Thresholds τH , τO , and τS define the feasible region. General Alignment applies in the interior, whereas Edge Alignment concerns behavior near edges and vertices where HHH constraints conflict.
Figure 4: The Edge Alignment process stack. Diagnostic signals first expose where general alignment is unreliable; structural capacity preserves conflicting objectives ( § 6.1 ); normative governance supplies legitimate plural content ( § 6.2 ); and uncertainty-aware interaction selects the appropriate process ( § 6.3 ).
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Table 2: Coverage of representative benchmarks. The three signal columns abbreviate the paper’s edge signals: structural conflict, normative disagreement, and epistemic ambiguity. A checkmark indicates direct evaluation as part of the benchmark’s primary scoring target.
Table 3: Composition of EdgeBench-Pilot .
Table 4: Process-aware scoring rubric for EdgeBench-Pilot .
Table 5: Human audit reliability for a 50-response stratified subset of EdgeBench-Pilot .
Figure 5: Lifecycle of Edge Alignment . Edge failures become actionable only when data collection preserves conflict, training objectives maintain value structure, evaluation reports process-aware diagnostic profiles, and deployment governance gives ambiguous or contested cases appropriate oversight.
Table 6: Lifecycle challenges for making edge failures visible and actionable.
Figure 6: Macro-level evidence of the paradigm imbalance in alignment research. The literature remains dominated by general alignment, while edge-alignment work has only recently begun to emerge.