This paper has been superseded
This paper has been consolidated into Context Management in Autoregressive Language Models (DOI: 10.5281/zenodo.19431948). The content below is retained for reference.
Executive Summary
Paper XX established that fixed reset-then-prime priming achieves CE 0.733. A natural question is whether adaptive priming — a "shepherd" agent that sees each probe and generates a probe-specific coordination signal — could do even better. This paper tests three shepherd strategies on Qwen 2.5-7B across 160 probes: storyteller (indirect narrative analogy), provocateur (challenge and reframe), and director (explicit instruction).
The results are unambiguous: adaptive priming is not just unnecessary, it is actively harmful. All bare shepherd strategies perform worse than the no-coordination baseline, with gap closures of -12.3% (storyteller), -26.7% (provocateur), and -37.1% (director). The more directive the shepherd, the worse the outcome. When reset is added before shepherd intervention, all three strategies converge to the performance of reset alone — the shepherd content contributes nothing beyond what the reset already provides.
The ratio tells the story clearly: reset accounts for 22:1 of benefit relative to the best shepherd content. The best protocol remains the one from Paper XX — reset followed by fixed domain priming (CE 0.733). No adaptive strategy tested in this paper improves upon it; inserting a storyteller layer between reset and domain priming degrades that result by 0.12 nats. The shepherd outputs are meta-commentary about the domain, not domain content: they activate narrative or instruction-following mode, and move layer-10 activations further from the expert than neutral priming does.
Key Findings
- All bare shepherds worse than baseline: Storyteller -12.3%, provocateur -26.7%, director -37.1% — more directive = more harmful
- Reset dominates: With reset, all shepherd strategies converge to reset-alone performance — shepherd content contributes nothing
- 22:1 reset-to-content ratio: Reset accounts for 22x more benefit than the best shepherd content
- Best condition unchanged: Reset + fixed priming (CE 0.733) from Paper XX remains best; adding a storyteller layer degrades it by 0.12 nats
- Meta-commentary, not domain content: Shepherd outputs move activations further from the expert than neutral priming (L2 28.1-28.5 vs 21.6)
Key References
- McEntire (2026) — Ritual Shape: reset then prime (Paper XX)
- McEntire (2026) — Ensemble Gravity: priming selection closes 48.9% of the coordination gap (Paper XIX)