The Anchor Protects What It Names: Selection Mechanics and Recall Collapse in Compact Context Compression for Autoregressive Language Models

Abstract & Summary Agent architectures that summarise, compact, or re-summarise their own working state are recursive rewriting systems, and recursive rewriting degrades the content it carries. The standard mitigation is to re-inject a compact anchor—a state summary, a fact list, a SYSTEM_NOTE—rather than the source itself. That mitigation is usually justified by a compression intuition: the anchor is a lossy encoding of the source, so protection degrades smoothly and everywhere as the anchor shrinks. We test the intuition directly and find it false. A compact anchor is not a compression; it is a checklist. Forty technical passages carrying 708 curated terms were rewritten recursively ten times under five conditions differing only in what context block precedes the current text: nothing (A); the full source (B); token-matched topic-irrelevant filler (C); a model-generated compact semantic signal held fixed (D); and the same signal regenerated at every step from the drifting text (E). The protocol was run three times: twice on Qwen2.5-7B-Instruct and once on Mistral-7B-Instruct-v0.3. Partitioning each passage's tracked terms by whether the compact signal names them separates condition D into two populations that behave like different experiments. On named terms the 73-token signal is statistically indistinguishable from the 183-token full source in all three runs ($D - B = -0.015, -0.013, +0.005$; all $p > 0.27$). On omitted terms it loses roughly one fifth of the full source's protection in all three runs ($-0.230, -0.210, -0.201$; all $p < 10^{-4}, d_z = -1.10, -0.90, -0.98$). The within-passage gap between the two populations is created by the anchor, not carried by the terms: a full source anchor compresses the gap ($+0.029, +0.080, +0.057$) while a compact signal amplifies it ($+0.244, +0.277, +0.263$), a difference-in-differences of $+0.215, +0.197, +0.206$ across two architectures and two independent runs ($d_z = 0.99, 0.80, 0.95$; all $p < 10^{-4}$). Refreshing the anchor destroys it. Condition E begins from D's signal verbatim and regenerates it from each step's own output; by generation 10 it retains none of D's advantage on named terms ($E - B\text{ named} = -0.120, -0.088, -0.093$) and is statistically indistinguishable from having no anchor at all. On Mistral it is significantly worse than no anchor ($E - A\text{ term retention} = -0.065, d_z = -0.70, p_{\text{Bonf}} = 0.0007$). The mechanism is directly observable: the regenerated signal's own coverage of the source decays monotonically—$0.686 \to 0.611, 0.657 \to 0.620, 0.487 \to 0.444$—because a projection of the current text cannot reintroduce what the current text has already lost. A self-refreshed anchor is a checklist that forgets the items. Two supporting results. Prompt length is not the mechanism: a token-matched irrelevant block captures none of the anchor's benefit in any run. And entailment measured in the precision direction—summary entailed by source, the direction most monitoring stacks implement—is blind to this failure, reading $0.93\text{--}0.96$ in chains that have discarded a fifth of their technical vocabulary. Drift under recursion is a recall failure with essentially no precision signature. We give a formal decay model that predicts these results from a single structural assumption (anchor support is binary and item-wise), and a control-plane design that follows from it: the anchor plane must be derived from the source once, at ingestion, and be immutable thereafter, and its coverage must be treated as the system's published fidelity specification rather than as a compression hyperparameter. The Checklist Law of Compact Anchors: Splitting tracked technical terms by whether a 73-token compact signal names them reveals two distinct populations. Terms explicitly named survive at full source strength ($D - B = -0.015, -0.013, +0.005$; all $p > 0.27$), while omitted terms lose roughly a fifth of full source protection ($D - B = -0.230, -0.210, -0.201$; all $p < 10^{-4}, d_z \le -0.90$). Amplification of the Semantic Gap: A full source anchor compresses the within-passage gap between named and omitted terms ($+0.029\text{ to }+0.080$), whereas a compact anchor amplifies it ($+0.244\text{ to }+0.277$)—a difference-in-differences of $+0.206$ ($p \le 10^{-4}$), directly contradicting lossy compression models. Destruction via Self-Refreshing: Regenerating the anchor at each step from the drifting text removes its benefit entirely ($E - B\text{ named} = -0.120\text{ to }-0.093$) because the signal's own coverage of the source decays monotonically ($0.686 \to 0.611$). The Fallacy of NLI Precision Dashboards: NLI faithfulness computed in the summary-entailed-by-source direction stays near $0.87\text{--}0.98$ in chains that have discarded 20% of their technical vocabulary. Iterative semantic drift is a pure recall failure with essentially no precision signature. Relation to Prior Work: This empirical study directly tests and updates the theoretical framework established in Guarding the Signal: A Framework for Identifying and Repairing Semantic Drift in Generative AI (Sweeney, 2025; Zenodo DOI: 10.5281/zenodo.15809538). While the 2025 framework correctly identified the necessity of symbolic state anchors for preventing recursive collapse, this deposit supersedes its diagnostic mechanics by proving that anchoring acts as strict selection rather than lossy compression, and that drift detection requires recall-direction instruments.