Dynamic Context Runtime: Bounded Attention over Unbounded History
https://cybersec.org.za/research.html #AI #cyber #security #dev #Africa #LLM #context #rot #cyber
https://cybersec.org.za/papers/dcr-bounded-attention.pdf
Dynamic Context Runtime: Bounded Attention over Unbounded History
Language models degrade as stale and superseded material piles up in their context. Making the window bigger does not fix it. This report describes a runtime that keeps history unbounded and attention bounded — storing everything as immutable spans and a typed provenance graph, then assembling a small working set each turn by solving a knapsack under an explicit token budget. It includes a zero-dependency Rust implementation and an ablation that names which mechanisms are actually carrying the result.
4.19M tokens of history
235 tokens per query
7 / 7 probes answered
17,835× less attention than the full transcript
48,651 state nodes held
28× history growth, flat working set
13,721 lines of Rust
0 external dependencies
152 tests
5 controls found that could not fail
0.4% of stored spans ever rendered to the model
47% cheaper with a mechanism the paper argues for switched off