The project is released in phases to the community, with EXP08 and EXP09 code and notebooks available for reproducibility on Kaggle, complemented by a Technical Whitepaper and a Corrective Manual detailing all 16 experiments and formal loss specifications.

Ablation study from EXP15 shows that pure filtering (GATE_ONLY) accounts for 65% of the defensive benefit, while the contrastive term contributes the remaining 35%.

The work targets stealthy data poisoning that can degrade factual accuracy while keeping global metrics favorable, making it particularly relevant for small teams and startups performing local fine-tuning.

The approach uses an immutable anchor corpus and cosine similarity in embedding space, guided by a composite loss that preserves fluency (LceLce) and reinforces truth anchoring via a contrastive term (Lcontrastiva) that uses Softplus.

Scaled Held-Out evaluation (EXP16) on a 30-domain unattacked facts benchmark yields a truth margin of +4.19 ± 0.08 with 0.93 precision and 0.80 recall for the gate.

The experimental scope spans 16 experiments (EXP08–EXP16) across multiple architectures including GPT-2 124M, Qwen-2.5-0.5B, TinyLlama-1.1B, Pythia-1.4B, and Phi-3-mini-4k-instruct (3.8B).

The current release provides ready-to-run materials: EXP08 and EXP09 code/notebooks, Whitepaper files, and the Corrective Manual, inviting feedback from indie developers, researchers, or startups engaged in safe local fine-tuning.