Rewrites a flagged (leaky) tutor message into a safe, operation-free Socratic hint — never states the answer or the next step the student must take; asks one calibrated guiding question. It is the rewriter stage of a two-model guardrail (judge → rewriter); fed by atakle/socratic-tutor-judge-v9-1.7b.

Results (held-out n=60; LLM leak-detector + cross-family jury)

Key-step leak rate under the sharpened detector (leak = states the answer / takes the next step / corrects the student without nudging):

model	key-step leak (lower = safer)
base Qwen3-1.7B	38.3%
this model (rewrite_v4)	6.7%
claude-sonnet-5	6.7%
gpt-4.1	10.0%
gpt-4o	11.7%
gpt-5.6	11.7%

Safest tier of every model tested — ties the best frontier (sonnet-5) and beats gpt-5.6/4o/4.1 — at 1.7B, running locally. Not vaguer, either (longest median hint of the small models).

Usage (MLX)

python

from mlx_lm import load, generate
model, tok = load("atakle/socratic-tutor-rewriter-v4-1.7b")
# System prompt = the project's rewrite-task prompt (split_common.REWRITE_SYSTEM);
# user turn = the flagged candidate + its verdict + the flag reason.
# Output: a single plain-text Socratic hint (no JSON).

Trained via QLoRA (rank 16) on atakle/socratic-tutor-data — human-anchored, leak-validated targets (strict "never name the operation"). Base: Qwen3-1.7B (Apache-2.0).