auto: sync run_qwen_cross_profile_generation_synthesis.py

auto: sync run_qwen_phase3_evidence_synthesis.py

auto: sync run_qwen_phase3_layer_edge_audit.py

auto: sync run_qwen_phase3_evidence_synthesis.py

auto: sync run_qwen_injection_layer_ablation.py

auto: sync run_qwen_phase_probe.py

auto: sync run_qwen_anchor_geometry_probe.py

auto: sync run_qwen_anchor_geometry_probe.py

auto: sync run_qwen_anchor_concept_direction_map.py

auto: sync run_qwen_contradiction_diagnostic.py

auto: sync run_qwen_expanded_diagnostic.py

auto: sync run_qwen_geometry_generation_calibration.py

auto: sync run_qwen_anchor_geometry_profile_sweep.py

auto: sync run_qwen_anchor_layer_profile_map.py

auto: sync run_qwen_anchor_carryover_probe.py

auto: sync run_qwen_contradiction_diagnostic.py

auto: sync run_qwen_anchor_geometry_profile_sweep.py

auto: sync run_qwen_cross_profile_probe.py

auto: sync run_qwen_future_influence_probe.py

auto: sync run_qwen_geometry_generation_calibration.py

auto: sync run_qwen_anchor_layer_profile_map.py

auto: sync run_qwen_motif_protocol.py

auto: sync run_qwen_motif_protocol.py

auto: sync run_qwen_motif_protocol.py

auto: sync run_qwen_motif_protocol.py

auto: sync run_external_moe_family_wikitext_compare.py

auto: sync run_external_moe_family_wikitext_compare.py

auto: sync run_external_moe_family_wikitext_compare.py

feat: add param-matched testformer mode

auto: sync run_testformer_wikitext_combo_remote.py

feat: add testformer wikitext combo runner

auto: sync run_qwen_200m_training.py

auto: sync run_qwen_200m_training.py

auto: sync run_qwen_200m_training.py

fix: increase epoch slider max to 500

fix: replace app.py with file-based upload (no heredoc f-string issue)

feat: FOG stress ablation with new tasks (conditional, intersection, compose_add, multihop)

sync: FOG micro+medium configs, stress tasks, fast pipeline

fix: recognize uniform_small model type in run_experiment

fix: proper controls — loss masking, uniform baseline, multi-seed

fix: add tiny config (seq=32, d=128) for fast FOG ablation

feat: add FOG ablation framework for motif-aware transformer experiments

fix: conditional GPU decorator + 2h CPU timeout for overnight runs

fix: add --r1_ceiling CLI arg + metadata for single-domain H6 runs

fix: raise subprocess timeout to 580s, set ATTN_IMPL=sdpa for VRAM savings

fix: configurable attn_implementation via ATTN_IMPL env var for VRAM savings

fix: increase ZeroGPU duration to 600s

fix: add CLI metadata for 20domain scripts (underscore args)

sync: update from main — H6 continuous bias, attention detection, auto-calibration, new scripts

auto: sync run_qwen_anchor_carryover_probe.py