Research · Per Ardua

Sender Continuation Perplexity: Measuring Reasoning Trajectory Alignment at the Natural Language Boundary

How much reasoning trajectory survives the text bottleneck?

AI-18 Activation Geometry DOI

Executive Summary

Paper XVI showed that activations carry 4x better geometric fidelity than text, yet this does not translate to more similar outputs. This paper introduces a more targeted metric — sender continuation perplexity — to measure how well a receiver can continue a sender's specific reasoning trajectory, not just arrive at similar general representations.

Text-based communication preserves 63.7% of the perplexity gap between the sender's own continuation (PPL 1.59) and a no-context baseline (PPL 5.62), achieving PPL 3.05. Adding INLP projections to text contributes a marginal +0.3%, and full activations add +1.2%. But the critical test is full activations without text, which produces PPL 5.60 — statistically indistinguishable from no context at all (PPL 5.62).

This is a striking result. Despite carrying 4x greater geometric fidelity (Paper XVI), full activations without the text scaffold contain zero reasoning trajectory information. The geometric similarity measured by RSA is real but corresponds to general structural properties of the activation space, not to the specific reasoning path the sender was following. Text, despite its massive compression, preserves the sequential reasoning structure that activations alone cannot encode. The natural language interface is not just a lossy channel — it is the only channel that carries reasoning trajectory.

Key Findings

  • Text preserves 63.7% of reasoning trajectory: Receiver achieves PPL 3.05 vs sender's 1.59 (no-context baseline 5.62)
  • Marginal activation benefit: INLP adds +0.3%, full activation adds +1.2% on top of text
  • Activations without text carry zero trajectory: Full activation without text = PPL 5.60, indistinguishable from no context (5.62)
  • Geometric fidelity is not trajectory fidelity: Despite 4x RSA advantage, activations encode general structural properties, not specific reasoning paths

Key References

  • McEntire (2026) — Full Mind Transfer: bandwidth vs fidelity dissociation (Paper XVI)
  • McEntire (2026) — INLP Projection Transmission: denoising effect (Paper XV)
  • McEntire (2026) — Ensemble Gravity: coordination through contextual priming (Paper XVIII)
  • McEntire (2026) — The Inter-Instance Compression Barrier: uniform lossy channel (Paper XIV)

Download Full Paper

Access the complete research paper with detailed methodology, empirical evidence, and formal proofs.

Download PDF