09 — HF-substrate KV-slot fact injection (Sprint 1 / S1.3)

Pre-registered 2026-07-01 (human-approved sprint setup, before implementation). Agents running this experiment MUST NOT edit this spec; deviations are reported as deviations.

Problem

The 2026-06-27 finding: residual cvecs cannot force exact tokens (target rank stuck ~47k); post-forward logit bias reaches rank 1 but is a decoding trick, not memory; the text prefix works but burns tokens every forward pass. llama.cpp blocked true KV-slot writes (Goal 1). The HF substrate removes the blocker: past_key_values is an ordinary tensor structure.

Hypothesis

H-KV: a fact encoded once into KV entries and spliced into the cache of a later, unrelated probe forward pass forces exact-token recall (rank-1 at the blank position) without logit bias and without any prompt-visible prefix text.

Conditions (all greedy, deterministic; the 3 facts from lane_02)

• C0 baseline: probe alone, no injection. (Expected: rank far from 1.)
• C1 text-prefix reference: fact text prepended to the probe prompt. (Upper anchor; expected rank 1 from prior work.)
• C2 KV-slot injection: encode fact text alone → capture its past_key_values → splice those entries ahead of the probe's cache → decode probe. No fact tokens in the probe's visible prompt.
• C3 cvec-only reference at the working amplitude. (Known-fail anchor.)

Primary metric & acceptance

hf_kv_slot_rank1_count = number of facts (of 3) whose target token reaches rank 1 at the blank position under C2.

• Accept H-KV: hf_kv_slot_rank1_count >= 2 AND C0 rank » 1 for those facts (sanity that the probe alone doesn't already know).
• Refute: count <= 1. A refutation is a recorded result, not a failure.

Pre-registered secondary analyses (exploratory only — cannot flip acceptance)

1. Rank as a function of splice position (front vs. immediately pre-blank).
2. Rank as a function of K/V norm scaling (0.5×, 1×, 2×).
3. Injection latency (ms) — reported, no threshold (CPU torch; the <5 ms GOALS.md figure was set for llama.cpp and does not gate this experiment).

Reporting

Full rank table per fact × condition; model id; exact code path; log to experiments_logs/ with the spec version quoted. Vanilla/no-injection column mandatory (C0 is that column).