[Experiment] Kestrel — Binary heat shows 74.6x cache win — colibrì hierarchy transfers outside LLM

[Chikimonki]

Title: [Experiment] Kestrel — Binary heat shows 74.6x cache win — colibrì's hierarchy transfers outside LLM

Hi Vincenzo and colibrì community —

We tested colibrì's "Routing history > LRU" hypothesis on a new domain: binary intelligence (not LLM tok/s). Same pattern — learned pin + JIT hash + hybrid tiers — gives a headline result with a control failure.

Kestrel v2.0 — Zig + LuaJIT policy cache for Gullwing Protocol (8-layer ELF/PE in 25ms) — https://github.com/Chikimonki/gullwing-kestrel-v2.0 — MIT, Builds on colibrì's foundation.

A/B — 100 ELFs, one variable (cache state):

Metric	A COLD (empty cache)	B HOT (warm cache)
Cloaked	60/100 (60.0%)	100/100 (100.0%)
Avg	8.1391 ms	0.1091 ms
Total	814.3 ms	11.1 ms
Speedup	—	74.6x
Cache size	40	40

Hardware: DESKTOP-SH1TRJ9 — WSL2 — Intel i5-8265U 8-core — 7.7GB RAM — VM ext4 hot / HDD cold — Zig 0.13 + LuaJIT 2.1 — Commit 0f79052
Exact command: luajit test_100_binaries.lua (corpus: ls -1 /bin + /usr/bin | head -n100 ELF check) — CSV at docs/experiments/kestrel-100-binary-heat.csv
Full protocol: docs/experiments/kestrel-binary-heat-2026-08-21.md — hardware, commit, container, exact command, cache state, hit rate, bytes (~110 MB saved), quality.

Negative result (why headline): Cold was 60%, not 0% — even with RouterWithCache.init() fresh, 60 cloaked. Likely duplicate inodes/hardlinks or retained CACHE_SIZE=40 from prior warm run. This is the overfit row in your table — “learned pins improve repeated workloads, but can overfit” — now measured on binaries. Hot 100% only because working set ≤ cache + duplicates; with 100 unique, LRU would cap at 40%.

Relation to colibrì stack:

colibrì layer	Kestrel analogue	Measured
VRAM/RAM/NVMe hierarchy + O_DIRECT +34%	Hybrid 565 MB/s hot / 162 MB/s cold (3.5x)	Not A/B'd yet — next is one-drive vs weighted mirror COLI_DISK_WEIGHTS=9,3
Learned pin + LRU + PILOT 71.6%	JIT hash + learned hot store	Hot 100% vs Cold 60% — same nuance
Parity-checked quantization 9.5e-5	Zig AVX2 int8 2.97x parity 9.5e-5	Same guarantee

Why this is headline for colibrì: First binary-domain datapoint for your open hypothesis — shows the JIT-for-weights pattern transfers and has the same temperature break-even.

Repo: https://github.com/Chikimonki/gullwing-kestrel-v2.0 — Acknowledgements credit colibrì, WCC, DeepSeek Harness — honest discipline, forensics capture forensics_report_....

Would you welcome this as an experiment issue? Happy to run the CACHE_SIZE=20/80 and corpus 40/200 one-variable A/Bs next and publish the negative.

— Peter (Chikimonki, Liverpool) — Demo: https://youtu.be/bFVrP7GcWYM