Methodology · Ikigai QoE

The rule underneath everything: no number in a deliverable originates from a language model. A model may label a deposit or draft a paragraph; it never produces a dollar amount, a ratio or a score. The report serializer rejects any figure whose provenance record says otherwise.

CrossTie — the four-way tie-out and the Cash Proof

CrossTie asks whether the revenue is real. It works at transaction level across four independent legs: the general ledger, the bank, the filed return, and the IRS transcript.

Classify every deposit — customer receipt, card batch, loan proceeds, owner contribution, intercompany transfer, refund reversal, asset sale, grant or relief, or unclassified. Classification uses a deterministic model over amount, timing, recurrence, counterparty and account role.
Adjusted deposits B* = all deposits less those classified as non-revenue.
Match B* to ledger revenue by amount within tolerance, date within five business days and counterparty, allowing many-to-one matches for card settlements.
Proof ratio ρ = matched revenue / reported ledger revenue.
Book-to-tax delta δ[y] = (ledger revenue − tax gross receipts) / tax gross receipts, per year.
Transcript delta ξ[y] = filed-return gross receipts − IRS transcript gross receipts. Any non-zero ξ is escalated to the top of the report and is never resolved automatically.
Residual — unmatched amounts are split into timing (reversing within 45 days across a period boundary) and unexplained.
Cash Proof — reported on exactly three windows: trailing twelve months, most recent fiscal year, prior fiscal year.

Proof ratio grading
Proof ratio ρ	Grade	Consequence
≥ 0.97	Verified	—
0.92 – 0.97	Substantially verified	—
0.85 – 0.92	Partially verified	Disclosure required
< 0.85	Unverified	The file cannot be certified. There is no override.

The Defensibility Index

Every candidate adjustment is scored 0–100 on five axes and tiered. The headline earnings figure includes Tiers 1 and 2 only.

DI = 0.30·E + 0.25·D + 0.20·M + 0.15·T + 0.10·C (each axis 0–100)

E — Evidence completeness. Linked evidence weighted by document strength and the share of the dollars it covers: contract or appraisal 1.0, third-party invoice 0.9, cancelled check 0.8, internal memo 0.4, assertion only 0.1. Capped at 100.
D — Discretion. Whether a new owner would actually stop the expense. Inverted for deferred maintenance, which a new owner must spend.
M — Magnitude normalcy. 100 − 30·|z|, where z is the amount as a share of revenue against the industry peer band. Floored at zero.
T — Timing consistency. Items claimed as non-recurring that appear in all three years score 15 or less.
C — Counterparty independence. 100 at arm's length; 30 for a disclosed related party with a third-party benchmark; 10 for an undisclosed affiliate.

Defensibility tiers
DI	Tier	Treatment
≥ 80	Tier 1 · accepted	In the headline
60 – 79	Tier 2 · supportable	In the headline
35 – 59	Tier 3 · contested	Excluded, disclosed with amount and reason
< 35	Tier 4 · rejected	Excluded, disclosed with amount and reason

All five categories the SOP names — non-recurring items, owner compensation against market, related-party transactions, deferred maintenance, and cash-versus-accrual differences — appear on every lender schedule, including any where nothing was identified, with the basis for that conclusion. Owner compensation is always the difference from a market replacement salary looked up by industry, metro area, revenue band and headcount; adding back the whole salary is treated as an error.

The Earnings Durability Score

A 0–1000 composite of six pillars, each normalized against an industry peer set, graded A–F. Each pillar reports its three largest drags so the score says what to fix.

EDS pillars and weights
Pillar	Weight	Built from
Revenue proof	25%	Proof ratio, book-to-tax delta, unexplained residual
Adjustment defensibility	20%	Dollar-weighted DI, contested share, add-back ratio
Concentration	15%	HHI, top-five share, retention, contract coverage, tenure
Margin stability	15%	Gross margin variance and slope, deseasonalized revenue volatility
Working capital	15%	NWC to revenue against peers, cash conversion cycle, volatility
Books and records	10%	Reconciliation lag, account count, journal-entry anomalies, closed-period edits, Benford screen

wDI         = Σ(amount × DI) / Σ(amount)                 over included adjustments
riskyShare  = Σ(amount where tier ≥ 3) / Σ(amount)
abRatio     = Σ(amount) / TTM revenue
defensibility = clamp(wDI×8.2 − riskyShare×520 − max(0, abRatio − 0.08)×1400, 0, 1000)
EDS         = Σ(weight × pillar)
grade       A ≥ 850 · B 700–849 · C 550–699 · D 400–549 · F < 400

The Benford screen reports a p-value. It is a prompt for questions, never an assertion of fraud.

The DSC Bridge

Adjusted, accrual-basis earnings carried to the coverage test the SOP requires. Historical earnings only — forward-looking inputs are refused on any SBA file.

CADS = (normalized earnings − maintenance capex) × (1 − t) + t × year-one interest
DS   = annual payment on the 7(a) tranche (≤ 120 months for the business portion)
       + real estate tranche + assumed debt + working-capital facility
       + seller note (zero while on full standby for the full term)
DSC  = CADS / DS        floor 1.25× (acquisitions, owner buyouts) · 1.15× (expansions)

loan = min( appraised value × (1 − injection),  price × (1 − injection),  $5,000,000 )
breakeven SDE = (floor × DS − t × interest) / (1 − t) + maintenance capex
headroom      = adjusted SDE − breakeven SDE

The bridge reports which constraint binds — value, price, the 7(a) ceiling, or coverage — the maximum supportable loan, coverage at +100bp and +200bp, the gap-fill options (price, buyer equity, standby seller note), and the adjustment headroom: the dollars of disallowed add-backs at which the deal breaks the floor. Deferred maintenance is netted into normalized earnings once; maintenance capex is the separate ongoing replacement cycle.

The Retrade Risk Model — scorecard regime

Version 1 is a transparent expert-weighted scorecard, not a fitted model. Its weights are set by judgment, every report says so in its header, and the regime is stamped into the run record. It does not carry a calibration statistic because it has not been fitted to outcomes.

multiple  = 2.6 + EDS/1000 × 1.4                       (credibility-adjusted)
EV        = adjusted SDE × multiple
P_retrade = clamp(0.98 − EDS/1000 × 0.95 + riskyShare × 0.90, 0.08, 0.95)
severity  = clamp(0.035 + riskyShare × 1.15 + (750 − EDS)/1000 × 0.10, 0.02, 0.35)
P_close   = clamp(0.94 − riskyShare × 1.40 − (800 − EDS)/1000 × 0.55, 0.25, 0.95)
proceeds  = (EV − EV × P_retrade × severity) × P_close

It graduates to a fitted model once 300 closed deals have captured outcomes. At that point the report will disclose the sample basis and the calibration statistics. Remediation items are ranked by how much each one changes risk-adjusted proceeds. Lender engagements carry no remediation plan.

The evidence ledger

Every figure carries a provenance record: source document, page and line, extraction confidence, the named transformation that produced it, and reviewer state. Records are append-only and hash-chained, so any alteration is detectable. In the portal, a figure in the report opens to the source line that supports it.

Human sign-off

A file reaches certified status only with a reviewer record: a CPA- or CVA-credentialed analyst, a timestamp and a signature. There is no automatic promotion and no administrative bypass. Every override of an engine determination is recorded with its reason.

The demo uses these formulas

The interactive demo computes the workbench, the Earnings Durability Score, the retrade scorecard and the DSC Bridge with the formulas above on a fictional file. Its five non-defensibility pillars are fixed values from a simulated engine run; its figures are illustrative.