ledger-models

name	ledger-models
version	2.0
type	autoresearch
description	Autonomous deal modeling with live market data
user-invocable	true
agent	LEDGER
agent_model	deepseek/deepseek-r1
mcps	["aegis","paperclip","filesystem"]
gstack_skills	["/benchmark"]
eval_metric	model_accuracy
eval_budget	120s
guard	sba_compliance = true

/ledger-models — Autonomous Financial Modeling

What it does

Builds and refines CPA acquisition financial models using AEGIS for heavy computation, Paperclip for live market data, and /benchmark for regression testing.

Protocol

Step 0: Load context

• memory/CFO_MEMORY.md — active deal models, assumptions
• dark-factory/VALUATION_*.md — existing valuations
• HENRY_BRAIN.md — deal criteria

Step 1: Fetch live market data (Paperclip)

1. SBA 7(a) current rates — Prime rate + spread, current terms
2. Comparable CPA sales — Recent transactions on APS.net, BizBuySell
3. Market multiples — Current CPA firm EBITDA multiples (accounting industry)
4. Interest rates — Federal Reserve, Treasury yields

Step 2: Build/update model (AEGIS → DeepSeek R1)

Route to AEGIS with DeepSeek R1 for quantitative work:

For each target deal:

INPUTS
  gross_revenue: $[from DD]
  asking_price: $[from listing]
  henry_buy_price: revenue * 0.4
  sba_loan: henry_buy_price * 0.9
  cash_injection: henry_buy_price * 0.1
  loan_term: 25 years
  rate: prime + 2.75% (current)
  
TRANSFORMATION MODEL
  current_labor_cost: revenue * 0.45 (industry avg)
  ai_replacement: labor_cost * 0.65
  labor_savings: current_labor_cost * ai_replacement
  post_ai_ebitda: revenue - (current_costs - labor_savings)
  ebitda_margin: post_ai_ebitda / revenue (target: 75-80%)
  
EXIT MODEL
  exit_multiple: 7x EBITDA (conservative)
  exit_value: post_ai_ebitda * exit_multiple
  total_roi: (exit_value - henry_buy_price) / henry_buy_price
  cash_on_cash_yr1: (post_ai_ebitda - annual_debt_service) / cash_injection
  payback_months: cash_injection / (monthly_ebitda - monthly_payment)

SENSITIVITY
  client_retention: [70%, 80%, 90%, 100%]
  ai_replacement: [50%, 60%, 70%, 80%]
  exit_multiple: [5x, 6x, 7x, 8x]

Step 3: Regression test (/benchmark)

Run /benchmark against previous model versions:

• Does the model still produce valid outputs for all 4 deal targets?
• Do the assumptions stay within HENRY standards?
• Are the SBA calculations compliant?

Step 4: Save results

• Write updated model to dark-factory/VALUATION_[ID].md
• Update memory/CFO_MEMORY.md with latest numbers
• Flag any deals where numbers changed significantly

Autoresearch Loop

• Metric: model_accuracy — how close model predictions are to actual deal outcomes (when available) or to industry benchmarks
• Budget: 120s eval
• Modify: assumptions, formulas, sensitivity ranges, data sources
• Guard: SBA compliance must always pass (loan limits, injection %, term limits)

Active Deal Models

Target	Revenue	HENRY Buy	SBA Loan	Monthly Pmt	Post-AI EBITDA	Exit 7x
TXS5345	$142K	$56.8K	$51.1K	~$470	$112K	$784K
TXS5513	$424K	$127-170K	~$144K	~$1,320	$254-297K	$1.78-2.1M
TXS5491	$910K	$364K	$327.6K	$3,020	$637K	$4.46M
TXS5450	$472K	~$189K	~$170K	~$1,560	~$330K	~$2.3M

MCP Integration Map

LEDGER (DeepSeek R1)
  ├── aegis: route quant work to DeepSeek R1
  ├── paperclip: fetch live SBA rates, comps, multiples
  ├── filesystem: read/write valuation files
  └── /benchmark: regression test models