Issues

• v2-7B forge: 50.0/44.5 vs base 62.2/53.7 — investigate ~20% relative drop after activation+pad+500 steps

  CambrianTech/sentinel-ai#165 · Filter by author joelteply opened on Apr 7, 2026
  1

• forge_v2_pipeline.sh stage 4: llama-cli hangs in interactive mode without -st flag

  CambrianTech/sentinel-ai#164 · Filter by author joelteply opened on Apr 7, 2026

• Hybrid attention defrag, Strategy A: skip non-full_attention layers (DeltaNet + quant only)

  CambrianTech/sentinel-ai#163 · Filter by author joelteply opened on Apr 7, 2026

• WSL2 ghost CUDA context: 1.6 GiB unrecoverable VRAM blocks 14B fp16 EvalPlus on 32 GB GPU

  CambrianTech/sentinel-ai#162 · Filter by author joelteply opened on Apr 7, 2026

• Defrag silently violates llama.cpp's unstated q_proj_out == hidden_size invariant; validation must include deployment-runtime load tests

  CambrianTech/sentinel-ai#160 · Filter by author joelteply opened on Apr 7, 2026

• Neuromodulation: extend MIMO forge controller from train-time to inference-time gain control

  CambrianTech/sentinel-ai#159 · Filter by author joelteply opened on Apr 7, 2026

• Predictive-coding importance metric: fifth column in the four-metric comparison

  CambrianTech/sentinel-ai#158 · Filter by author joelteply opened on Apr 7, 2026

• Sleep-phase consolidation: replay alloy stages into model weights (CLS / hippocampal replay)

  CambrianTech/sentinel-ai#157 · Filter by author joelteply opened on Apr 7, 2026

• Plasticity invariant: per-link weight-delta attestation with personality probe

  CambrianTech/sentinel-ai#156 · Filter by author joelteply opened on Apr 7, 2026

• Defrag breaks on hybrid attention models (Qwen3.5: linear_attention + full_attention)

  bug
  CambrianTech/sentinel-ai#154 · Filter by author joelteply opened on Apr 7, 2026

• Defrag validation harness — verify tensor surgery preserves model integrity

  enhancement
  CambrianTech/sentinel-ai#153 · Filter by author joelteply opened on Apr 7, 2026

• Crowdsourced forge telemetry: meta-learner trains on grid-wide surgery/recovery data

  CambrianTech/sentinel-ai#151 · Filter by author joelteply opened on Apr 3, 2026

• Recovery strategy selector in forge pipeline (ransac, gradient, hybrid)

  CambrianTech/sentinel-ai#150 · Filter by author joelteply opened on Apr 3, 2026

• RANSAC-style weight recovery after model surgery

  CambrianTech/sentinel-ai#149 · Filter by author joelteply opened on Apr 3, 2026

• Automatic disk/cache management for forge nodes

  CambrianTech/sentinel-ai#148 · Filter by author joelteply opened on Apr 2, 2026

• Vision model forging: LoRA on VL architecture

  CambrianTech/sentinel-ai#147 · Filter by author joelteply opened on Apr 2, 2026

• Automated forge queue: submit next alloy when current job finishes

  CambrianTech/sentinel-ai#146 · Filter by author joelteply opened on Apr 2, 2026

• MoE expert pruning: domain-specialized expert selection

  CambrianTech/sentinel-ai#145 · Filter by author joelteply opened on Apr 2, 2026

• Variable quantization: per-layer precision from pruning importance

  CambrianTech/sentinel-ai#144 · Filter by author joelteply opened on Apr 2, 2026

• Pruning after context-extend: warmup required or heads misidentified

  bug
  CambrianTech/sentinel-ai#143 · Filter by author joelteply opened on Apr 2, 2026

• Vision resolution extension via position embedding interpolation

  enhancement
  CambrianTech/sentinel-ai#142 · Filter by author joelteply opened on Apr 2, 2026

• Candle Qwen3.5-VL: visual encoder + projection for native persona vision

  enhancement
  CambrianTech/sentinel-ai#141 · Filter by author joelteply opened on Apr 2, 2026

• Run benchmarks through Candle/continuum inference — not Python/HF

  enhancement
  CambrianTech/sentinel-ai#140 · Filter by author joelteply opened on Apr 2, 2026

• Candle Rust: implement causal conv1d for Qwen3.5 Mamba blocks

  enhancement
  CambrianTech/sentinel-ai#139 · Filter by author joelteply opened on Apr 2, 2026

• Eval: write results incrementally, support resume from checkpoint

  enhancement
  CambrianTech/sentinel-ai#138 · Filter by author joelteply opened on Apr 2, 2026