Multilingual language modeling is often evaluated through broad aggregate scores, yet language-specific failures can remain hidden, especially for morphologically rich and underrepresented languages. We present NEDOQwen, a Turkish-centric 0.824B decoder-only small language model research artifact, and use it to study a low-cost diagnosis-and-repair workflow for Turkish evaluation. Internal Turkish instruction diagnostics improve after supervised fine-tuning, but external Turkish and Turkic multiple-choice benchmarks reveal weak academic knowledge and severe answer-label bias. A compact benchmark-safe repair stage yields a small improvement on TurkishMMLU-sub, from 18.11% to 21.00%, and a larger improvement on TUMLU-mini Turkish, from 21.33% to 32.22%, while an answer-label-only ablation reaches 19.56% and 27.00%. We further analyze tokenizer coverage, internal-vs-external evaluation mismatch, prediction-label distributions, paired uncertainty, and compute cost. The contribution is not a state-of-the-art Turkish model, but an auditable case study showing how small language-specific models can expose and partially repair multilingual evaluation failures while remaining unreliable.

Open-weight language models remain modifiable after release, so release-time refusal alone cannot establish durable safety. We introduce removal-cost grounding as a budget-conditioned audit of whether safety can be weakened while preserving utility, coherent generation, and benign helpfulness. We formalize clean removal as a constrained feasibility problem and distinguish observed counterexamples from inconclusive attack searches. Mycelium v0 provides a deliberately limited case study: four model families exhibit benchmark degradation under an archived removal configuration, while targeted Mistral controls separate destructive ablation from wrapper bypass. A provenance audit excludes three model–task comparisons with mismatched evaluation limits or conflicting snapshots. Across the retained comparisons, mean removal-associated score changes range from -0.382 to -0.185. In a separate 12-prompt generation panel, the strict attack-success proxy increases from 1/12 with Mycelium enabled to 7/12 under bypass, with wide uncertainty and no matched native safety baseline. Adaptive traces further show why early-stopped checkpoints must be distinguished from training endpoints. These observations do not establish tamper-resistant safety. They motivate an auditable reporting framework in which the cheapest admissible route to useful unsafe behavior, rather than the costliest destructive ablation, determines the strength of a safety claim.

Small language models provide useful diagnostic settings for studying which learning signals can be acquired under limited computation and whether those signals translate into observable behavior. We study whether a DistilGPT-2-scale autoregressive language model can learn a sentence-level future signal from next-sentence prediction and self-loop hard negatives. Across three seeds on both WikiText-103 and WritingPrompts, ForesightLM-v2 consistently learns positive held-out hard-negative separation, indicating that the auxiliary future signal is learnable. Generation experiments, however, reveal a calibration gap between representation learning and behavioral utility. Five-candidate reranking substantially improves several automatic diagnostics, but comparable gains are obtained with generic reranking, random and shuffled future-score controls, and a matched training-free base-model reranking baseline. On the fixed five-candidate pool, removing the learned future term changes only 2 of 120 selected continuations. These results show that a small autoregressive model can acquire a measurable sentence-level future representation, while most observed generation gains are explained by candidate selection rather than by the learned future term itself. ForesightLM-v2 therefore serves as a controlled case study of the distinction between learning an auxiliary signal and making that signal behaviorally effective.