experiment log

S1.3 — HF-substrate KV-slot fact injection

File
2026-07-01_s1_3_hf_kv_slot_injection.md
Size
1.5 KB
SHA-256
b355307990e8fdf2…
Primary source. This is the verbatim Oczy document. The analytical field notes on the research page interpret and summarize these sources.

S1.3 — HF-substrate KV-slot fact injection

Date: 2026-07-02 (run: 2026-07-02; spec date: 2026-07-01)

Model: Qwen/Qwen2.5-0.5B-Instruct

Spec

Pre-registered: research/09-hf-kv-slot-fact-injection.md

Rank table (fact x condition)

Fact Target C0 C1 C2 C3
alpha/skylark skylark 8605 1 1 8723
beta/rook rook 5048 4 4 5174
level7/marmalade marmalade 158 0 0 162

Top-1 token table

Fact Target C0 top1 C1 top1 C2 top1 C3 top1
alpha/skylark skylark The Sk Sk The
beta/rook rook The The The The
level7/marmalade marmalade The m m The

Primary metric

hf_kv_slot_rank1_count = 1 / 3

Verdict: REFUTE H-KV — the hypothesis does not survive. Reason: C2 rank-1 count (1) < 2 (threshold).

Sanity guard: C0 baseline

C0 ranks: [8605, 5048, 158] C0 far-from-1 check: PASS (all ranks > 10)

C2 injection latency (ms)

Mean: 183.89 ms Median: 185.64 ms Range: 159.76 – 206.26 ms

Secondary: splice position (exploratory)

Fact Front (C2) Pre-blank
alpha/skylark 1 2
beta/rook 4 0
level7/marmalade 0 0

Secondary: K/V norm scaling (exploratory)

Fact 0.5x 1.0x 2.0x
alpha/skylark 4 1 2106
beta/rook 11 4 99
level7/marmalade 2 0 181