Context Fences: Two-Phase Coordination Through Mode Switching and Domain Activation

Executive Summary

Papers XIX–XXII established that the effective coordination protocol is reset, then prime, then deliver. But why does the reset work? The throughline interpretation framed it as garbage collection: the reset clears residual activation biases, bringing activations back to a neutral baseline, with the optimal clearing signal approaching zero content. This paper tests that hypothesis on Qwen 2.5-7B with nine clearing strategies across four experiments, against an alternative: mode switching, where the reset installs a processing mode boundary rather than clearing accumulated state.

The garbage collector hypothesis is falsified on three counts. First, activation distance to neutral after clearing does not predict cross-entropy (Spearman rho 0.100) — if garbage collection were the mechanism, clearing closer to neutral should produce better results. Second, more tokens in the clearing instruction produce better results (rho -0.672), the opposite of the zero-content prediction. Third, reset benefit is constant at approximately 0.076 nats across 0 to 384 tokens of prior context — garbage collection would predict increasing benefit with longer contexts.

The correct mechanism is mode switching. A standard 18-token reset, which names both what to ignore and what to attend to, achieves CE 0.503 (best), while silence (four newlines, no mode signal) yields CE 0.699 (worst). The two stages are separable: across an 18-cell clearing-by-priming matrix, post-prime activation distance to the expert predicts CE (Spearman rho 0.858), and mode and domain act as orthogonal dimensions of the processing state. The resulting protocol is: install a context fence, then activate the domain, then deliver the task.

Key Findings

• Garbage collection falsified (1): Distance to neutral does not predict CE (rho 0.100)
• Garbage collection falsified (2): More clearing-instruction tokens = better results (rho -0.672), opposite of garbage collection prediction
• Garbage collection falsified (3): Reset benefit constant ~0.076 nats regardless of context length
• Mode switching supported: Reset installs a processing mode boundary, not state clearing
• Standard reset performance: 18-token reset = CE 0.503; silence = CE 0.699
• Mode and domain separable: Post-prime distance to expert predicts CE (rho 0.858); the stages compose independently

Key References

• McEntire (2026) — Ritual Shape: reset then prime; within-session repetition degrades (Paper XX)
• McEntire (2026) — Shepherd Agents: content between reset and prime degrades by 0.118 nats (Paper XXII)
• McEntire (2026) — The Coordination Problem Is Interference: the garbage-collector interpretation (Throughline II, Papers XVI–XXII)
• McEntire (2026) — Sender Continuation Perplexity: text transmits 63.7% of reasoning trajectory (Paper XVIII)