⚡ WHAT WOULD TRIGGER GERMINATION ON FORGE
Primary trigger: LAAM silent failures > 30 in 90 days — the rain cheque pattern
Each silent parse failure plants a potential ghost vector. 30+ in 90 days = coherence degradation that crosses the germination threshold.
Secondary trigger: VRAM saturation event lasting >10min → LAAM forced to skip validation → failures spike
Tertiary trigger: qdrant + redis diverge for >24h (split-brain) → RAG queries return stale data at scale
germination_score = Σ(L5_seeds × velocity × coherence) / γ₁
current: seeds=2 × velocity=0.02/week × coherence=0.12 = 0.12
trigger path: if velocity reaches 0.4/week → score crosses 0.70 → class action threshold
#1 · pelego threshold calibration
PREVENTS → L3 VRAM CASCADE
pelego dps=64 miscalibration means it triggers more often than expected. Calibrating sigma_gate=0.5 threshold prevents pelego + alphastar co-firing, which is the primary VRAM contention cascade path.
Priority: HIGH · Effort: LOW · Impact: BLOCKS L3
#2 · LAAM parse error logging
PREVENTS → SILENT L2→L5 GRADUATION
LAAM ingest must surface all parse failures as errors, not silent skips. This is the single most important fix on forge. Without visibility into parse failures, the germination score is invisible until it's too late.
Priority: CRITICAL · Effort: LOW · Impact: REMOVES BLIND SPOT
#3 · qdrant dimension validation on ingest
PREVENTS → GHOST VECTOR L4
Validate dim=768 on every ingest before write. Reject malformed vectors at the gate. This prevents the ghost vector L4 composition bug from forming. Simple guard, eliminates the highest-risk L4 path.
Priority: HIGH · Effort: LOW · Impact: CLOSES L4 PATH
KILL SHOT · forge harvest
"The fastest path to an L5 on forge is a silent LAAM parse failure that plants a ghost vector and gets queried at scale. The three prune items above close that path. Fix them before the velocity reaches 0.4/week."