Zihua Wu (NVIDIA) and Georg Gartner (TU Wien’s cartography division) argue that context is not a passive container but a structured spatial field, and that governing it is architecturally necessary rather than a prompt-engineering nicety. Their target is the assumption that a bigger window is a better one: the lost-in-the-middle effect, “lost in the distance” degradation as token distance grows, attention entropy growth, append-only KV caches that cannot retract a stale belief, and effective context utilization that collapses long before the advertised limit is reached.

The framework has three moving parts.

  • Three epistemic zones. Black fog (unobserved), gray fog (stored memory), visible field (the active reasoning surface), connected by four transitions: sense (B→G), recall (G→V), evict (V→G), expire (G→B). Each zone has a failure mode — hallucination, drift and bloat, overload and dilution — and each transition has composition failures too. The nastiest one for long sessions: destructive compaction that skips archival sends the original context to black fog permanently.
  • Seven operators, carried over from cartographic generalization theory: reconnaissance (what to explore), selection (what crosses a boundary), simplification, aggregation, projection (including compaction), displacement (repositioning for salience), and layering (typed namespaces). Selection, aggregation, projection and layering are structural isomorphisms to mapmaking operators; simplification and displacement are analogies.
  • An invariant triad — semantic, structural, salience — that no single transformation can preserve at once, exactly as no map projection preserves angles, areas and distances simultaneously. That mutual exclusivity is the argument for keeping the operators separate: it tells you which invariant a failed compression violated.

The empirical hook is a scored audit of four contemporary systems on a five-criterion rubric (Present / Explicit / Configurable / Automated / Documented): MemOS 3.29, Claude Code 3.00, Letta 2.86, OpenViking 2.71, for a mean of 2.96 out of 5 — roughly 60% of the operator space. Layering scores 5.00 across all four; projection averages 4.25; displacement scores 1.25, the lowest of the seven, despite addressing the best-documented failure mode in the literature. Where positioning happens at all, it is an accident of append-only ordering rather than a designed mechanism.

OpenViking’s own profile is the most useful line here for anyone running it: path-based selection, L0/L1/L2 tiered projection and filesystem namespacing each score 5, while simplification, aggregation and displacement each score 1. The predicted failure modes follow directly — noisy tool output contaminating the visible field, redundant retrievals diluting attention, and constraints drifting as context grows.

Take the caveats at face value. The framework is interpretive, not yet a full mathematical theory; the case studies are post-hoc, because the operators were partly induced from those systems, so “industry convergence” demonstrates descriptive adequacy rather than predictive power; and the five testable predictions plus the proposed operator-ablation benchmark are explicitly unvalidated. What is mechanically verified is structural — five Lean 4 theorems in a public repository, covering unique zone membership, boundary-operator uniqueness, non-commutativity of composition, and the information loss of destructive compaction.