QEvict recoverable quantized KV eviction evaluation
Status: bounded contract and witness for issue #6258. This is not a model-quality or production-throughput claim.
Named research object
This leaf evaluates QEvict: Recoverable Quantized KV Eviction for
Attention-Drift-Robust Long-Context Decoding, Garg et al.,
arXiv:2608.05326v1, published 2026-08-05.
The pinned paper PDF used for review was 16,507,328 bytes with SHA-256
baad72aecfec9adf0301aafb82c6a81ecdb4fb90b42a1ea2b15c5cdcc1f9b5dd.
The recipe identity in the contract is qevict/arxiv-2608.05326v1; no generic
“recoverable cache” surrogate is substituted.
QEvict’s named mechanism is a three-tier cache: important windows remain full precision, intermediate windows remain recoverable in quantized form, and the lowest-confidence windows are deleted. A later-important recoverable window is read, dequantized, and promoted. The paper names Future Missed Mass and Global LIR as drift diagnostics. The leaf measures the directly trace-replayable subset: future attention mass lost by ordinary irreversible eviction versus deletion by QEvict, recovery events/read bytes, and capacity. It does not pretend that the bounded trace reproduces Global LIR or downstream benchmark quality.
Contract
internal/qevicteval accepts only qevicteval/v1, the exact recipe pin above, a
content-addressed trace, and qevicteval/trace-replay-v1. Results are typed:
supported: pinned recipe and runtime were evaluated; decision isintegrateorabstainfrom the bounded evidence.unsupported: unknown contract/recipe, invalid artifact provenance, or invalid trace.delegate: the runtime is unknown or an independently captured runtime observation is absent. There is no silent fallback to modeled latency.
The result keeps capacity/drift evidence and latency evidence separate. Trace-derived bytes and missed mass are modeled. Latencies are observed only when platform, device, command, capture time, and evidence digest are pinned.
Reproducible witness
The three independent JSON fixtures under internal/qevicteval/testdata cover recovery,
no recovery benefit, and runtime delegation. The recovery fixture contains a historical
window ordinary eviction deletes but QEvict preserves at 25 bytes and later recovers.
Its deterministic replay reports:
| Metric | Ordinary eviction | QEvict | Evidence |
|---|---|---|---|
| Peak bytes in fixture accounting | 100 | 150 | modeled from fixture |
| Future missed attention mass | 0.41 | 0.01 | modeled from fixture |
| Recovery events / bytes read | 0 / 0 | 1 / 25 | modeled from fixture |
| Replay loop median (5 runs, 100,000 iterations) | 3.302 ns/op | 4.349 ns/op | observed, WSL Linux amd64, AMD Ryzen 9 9950X CPU |
| Recovery-copy median | n/a | 0.6407 ns/op | observed, same envelope |
Command:
go test ./internal/qevicteval -run '^$' \
-bench 'Benchmark(OrdinaryEviction|QEvictReplay|RecoveryRead)$' \
-benchtime=100000x -count=5
The captured output is committed as internal/qevicteval/testdata/wsl-benchmark.txt; its SHA-256 f7506bea881e438501367ca73f1fbce8d8dc83f9f9832e4e8527068d1a480779 is pinned by the observed fixtures. These nanobenchmarks measure
the Go trace adjudicator and a 25-byte recovery copy, not attention kernels,
dequantization, GPU memory traffic, token throughput, or model quality. Therefore the
observed runtime envelope supports the contract path only. Production integration remains
an abstain/delegate until a pinned implementation artifact and model/hardware witness exist.
No GPU evidence is required to establish this deliberately CPU-only contract witness; any
future CUDA claim must use the sanctioned lab path in docs/private-comms-channel.md and
must replace neither these provenance pins nor the modeled/observed distinction.
Decision
integrate means only that this exact bounded trace contains a recoverable future-attention
benefit and a complete runtime observation. abstain is returned when no recovery benefit
is present. Unknown or unobserved environments delegate. No universal quantization winner,
quality gain, or production performance gain is claimed.