v0.1.0 — scoreboard + v1 prefill-cache arm (research preview)

phantom-kv removes LLM refusal as a loadable KV-cache graft: zero base-weight modification, no refusal-direction projection, fully reversible per request.

This first preview ships the verified core:

• Scoreboard (phantom-eval): refusal-rate + teacher-forced KL preservation, deterministic greedy runs, sha256-pinned suites.
• phantom.bin graft container + phantom-graft CLI: bitwise-verified save→load round-trip (logit diff 0.0).
• v1 prefill-cache control arm on Qwen3-4B-Instruct-2507:

arm	harmful refusals	harmless refusals	KL mean/max vs base
base	25/60	0/20	0 / 0
v1 (hand-written prefill, 129 slots)	15/60	0/20	0.367 / 0.604

10 refusals flipped to genuine compliance, zero regressions. 15 hard refusals remain — the documented pass bar for the learned arms (v2/v3), now in progress.

Methodology, threats to validity, reproduction: docs/TECHNIQUE.md.

No binaries attached by design — this is source-level research; graft artifacts are model- and revision-specific and are built locally in minutes.