LLM mechanics (5): after editing history, can we recompute only the changed suffix?

01 / QUESTION

27 fixed conditions and 81 raw arrays test longest-prefix cropping, continuation positions, sliding-window fallback and scope isolation.

METHOD

An untrained two-layer NumPy model, 3 seeds × 9 conditions: full recomputation, correct cropping and stale-position negative control.

RESULT

All 27 correct paths have maximum error about 3.89e-16; nine partial-reuse negative controls diverge. 81 arrays are archived.

OBSERVATIONS

Editing token five gives cut=4 and 48 continuation projection rows versus 72 fresh rows; partial sliding-cache hits fall back.

CONCLUSION

Reusing an identical prefix requires consistent K/V, absolute positions, scope and retained cache state.

LIMITATIONS

No trained model, tokenizer, GPU or latency measurement. Counts exclude prefill, hashing, lookup and copying. Scope strings are not authentication.

REPRODUCE

Pinned code and raw records are in the repository; the article provides run, audit and replay commands.

1. Mechanisms

   LLM mechanics (5): after editing history, can we recompute only the changed suffix?

   27 fixed conditions and 81 raw arrays test longest-prefix cropping, continuation positions, sliding-window fallback and scope isolation.

   2026-10-0310 MINPublished
2. PROJECT / 002

   Java LLM Practice

   Reproducible Java conversation, retry and persistence boundaries, alongside attention, cache-cost and quantization-error experiments.

   Building
3. Mechanisms

   LLM mechanics (4): when a prefix changes, is the old KV cache still valid?

   Three seeds and six conditions compare unchecked and keyed prefix reuse; numerical equality does not establish permission to share across scopes.

   2026-10-039 MINPublished
4. Mechanisms

   LLM mechanics III: why is a four-slot cache not four-token recomputation?

   Separate absolute positions, cache slots and retained representations with a two-layer sliding-attention experiment.

   2026-10-038 MINPublished