The Math Behind Monce Suite

Charles Dana — Monce SAS, April 2026

1. Architecture

Monce Suite is an orchestrator that sits in front of 10 independent Snake SAT classifiers. Each classifier runs on its own EC2 instance with no LLM dependency in the default path.

  User text
      │
      ▼
 ┌──────────┐
 │  Monce   │ POST /monce-haiku or /monce-sonnet
 │  Suite   │ ─── fan-out to 10 /comprendre ──▶ all with anthropic=False
 └────┬─────┘
      │
  ┌───▼────┐
  │ Trust  │  aggregate trust score from 10 responses
  │ Assess │  high (≥65): return deterministic
  └───┬────┘  low (<65): one LLM call to synthesize
      │
  ┌───▼────┐
  │ Claude │  dynamic prompt built from what the
  │ (1x)   │  deterministic layer actually found
  └────────┘

2. The Economics

Without the suite, each service with anthropic=True makes its own LLM call. 10 services = 10 LLM calls per user input. With the suite:

Best case: 0 LLM calls. All 10 classifiers return high trust. Cost: $0.
Worst case: 1 LLM call. Trust is low, Claude synthesizes. Cost: 1x Haiku or 1x Sonnet.

The expected cost is 0.3x of a single LLM call (70% of inputs get deterministic-only).

3. Trust Score Assessment

Each downstream classifier returns a trust score (0-100) measuring extraction quality, signal density, classification confidence, and model agreement.

Zone	Threshold	Action
High trust	≥ 65	Return deterministic. No LLM.
Grey zone	35-64	LLM enhances weak areas.
Low trust	< 35	LLM full synthesis.
Too few OK	< 3 services	LLM with whatever is available.

4. Dynamic LLM Prompt

The LLM prompt is not static. It is constructed dynamically from whatever the 10 classifiers actually returned. Each classifier's highlights are extracted programmatically:

Predictions with confidence levels
Product matches and compositions
Client identifications
Anomaly detections
Routing decisions

The LLM acts as a synthesizer and tie-breaker, not a re-classifier. It never re-analyzes the raw text from scratch — it works from the structured outputs of 10 specialized models.

5. The 10 Classifiers

Service	Domain	Models	Best Metric
emailclassifier	Email triage	10 Snake	0.958 AUROC
quoteclassifier	Quote analysis	10 Snake	0.894 AUROC
requestclassifier	Complaint routing	4 Snake	0.990 AUROC
furnitureclassifier	Field sales scoring	1 Snake (27 feat)	0.860 AUROC
businessclassifier	Task dispatch	1 Snake (8 feat)	0.726 AUROC
selfservice	Product search	10 Snake	99.4% typo acc
salesagentclassifier	Prospect scoring	5 Snake	93.3% acc
negociationclassifier	Offer analysis	5 Snake	0.936 AUROC
procurementclassifier	Invoice control	2 Snake	98.6% acc
benchmarkclassifier	Supplier benchmark	10 Snake	0.959 avg AUROC

6. Theoretical Foundation

All classifiers use Snake SAT v5.2.1, based on the Dana Theorem (2024): any indicator function over a finite discrete domain can be encoded as a SAT instance in polynomial time. Training complexity: O(L × n × m × b). Inference: O(L × clauses). No backpropagation, no neural networks.

7. Latency Budget

Phase	Time	Cost
Fan-out to 10 services	50-300ms	$0 (CPU only)
Trust assessment	<1ms	$0
LLM fallback (if needed)	1-4s	1x Haiku or Sonnet
Total (deterministic)	50-300ms	$0
Total (LLM-enhanced)	1.5-4.5s	1x call