anima-agent-lane-tooluse-corpus

Agent-lane tool-USE demo corpus for anima's rung-0 tool-use grounding fire (design §6/§7). It teaches the byte-LM mouth the sentinel tool-call grammar (0xFE fact_lookup KEY 0xFF, vocab256-safe dead-UTF-8 delimiters) + grounding behaviour — NOT tool facts/trivia. Deterministic, $0.

two register variants (the key finding)

• agent_lane_chatreg.txt (1,060,800 B) — grammar taught IN the 사용자:/도우미: chat register. This is the variant that delivered the 🟢 F-TOOLUSE-FABDROP terminal PASS (fabrication 0.5556 → 0.0).
• agent_lane_5lang_disjoint.txt (980,880 B) — grammar in a plain-prose register DISJOINT from the chat turn. Closed-negative: the grammar was learned but stayed siloed and did NOT transfer to the chat surface (FABDROP FAIL). Lesson: teach the grammar in the same register it will fire in.

shape distribution (each variant)

a/b/c/d balanced 1200 each (needs-tool · no-tool-needed · don't-know-call · tier-too-low-refuse). Hard invariants (re-verified): fabricated_result_count == 0, sentinels balanced (0xFE=0xFF=3600), all non-frame bytes valid UTF-8, philosophy grep ([role:/[persona:/ [character:/[assistant:/[system:) == 0. Held-out probe values appear in NEITHER variant (falsifier leak guard).

philosophy (p1..p8 — HELD)

The 0xFE/0xFF are LEARNED grammar bytes, NOT identity injection. NO system prompt / persona / role / RLHF. sha256 manifest: SHA256SUMS.txt.

Scope (a_scale_honest_scope): TOY 18M rung corpus; mid/7B transfer UNVERIFIED.