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

The Anchor Protects What It Names 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, on the intuition that the anchor is a lossy encoding whose protection degrades smoothly and everywhere. This work tests that intuition and finds 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 the context block preceding the current text: nothing, the full source, token-matched irrelevant filler, a compact model-generated signal held fixed, and the same signal regenerated at every step from the drifting text. The protocol was run four times across two inference engines and two model families. Partitioning each passage's tracked terms by whether the compact signal names them splits the compact-anchor condition into two populations. On named terms a 73-token signal is statistically equivalent to the 183-token full source (two one-sided tests at ±0.05, established in every run). On omitted terms it retains roughly a fifth less than the full source and performs like no anchor at all. A randomised, yoked assignment — in which each term appears named in one arm and omitted in its complement, at the same token budget and under the same template — confirms that the effect is caused by naming rather than by any property of the terms (+0.143 pooled, p = 1e-12). Coverage does not buy protection proportionally: an anchor naming 20% of the terms is inert, while one naming 80% of them, at 57 tokens, is statistically equivalent to re-supplying the entire 183-token source. Refreshing the anchor from the model's own output destroys it. A self-refreshed anchor is statistically equivalent to having no anchor at all on the very content it began by naming, and the mechanism is directly observable: the regenerated anchor's own coverage of the source decays, because a projection of the current text cannot reintroduce what the text has already lost. Two further results qualify the finding. Naming a term causes the term to survive but does not detectably preserve the proposition it belongs to, so the protection reaches items rather than the relations between them. And entailment measured in the precision direction — summary entailed by source, the direction most monitoring stacks implement — is blind to this failure, reading 0.93–0.96 in chains that have discarded a fifth of their technical vocabulary. Drift under recursion is a recall failure with essentially no precision signature. The deposit contains the manuscript and its LaTeX source, all generated texts and per-step metrics for every run, the experiment harness, and analysis scripts that regenerate every reported number.