Vibecheck Deep Dive — Confidence-Gated Cascade

Architecture

This is the meta configuration repo for the Vibecheck Deep Dive cascade. It contains no model weights — weights are stored in the component repos below.

Component Repos

Component	Repo	Task
Quick Vibe (Tier 1)	`itsLu/mentalbert-v5-source-aware`	Cooperative source-aware 8-class MentalBERT
Pathway A (Tier 2)	`itsLu/mentalbert-v5-specialist-a`	CSSRS-pure 6-class specialist
Pathway B (Tier 2)	`itsLu/longformer-v5-crisis-evidence`	Binary crisis-evidence Longformer

Routing Logic

ROUTE_TO_DEEP_DIVE = (
    top1_prob < 0.65
    or (top1_class in {Depression, Suicidal} and margin < 0.2)
)
# Pass-through: Normal and Directed Aggression always use Quick Vibe

Fusion Gate

crisis_boost = alpha * pathway_b_crisis_prob  # alpha=1.5
fused_logits[Suicidal]   += crisis_boost
fused_logits[Depression] -= 0.5 * crisis_boost  # mass conservation
final = softmax(fused_logits)

Load Pattern

import json
from huggingface_hub import hf_hub_download

cascade_cfg = json.load(open(hf_hub_download('itsLu/vibecheck-deepdive-cascade', 'cascade_config.json')))
pathway_a   = load_specialist_a(cascade_cfg['pathway_a_repo'])
pathway_b   = load_specialist_b(cascade_cfg['pathway_b_repo'])
result = cascade_predict(text, quick_vibe, pathway_a, pathway_b, cascade_cfg)

Test Metrics

Metric	Value
Accuracy	84.82%
F1 Macro	0.8558
Total bleed	1179
Dep→Sui	482
Sui→Dep (critical)	697
% routed to Deep Dive	12.2%