Seven design errors, and not one of them is an engineering error
A scorecard nobody owns. A limit table that hands the same figure to a firm ten times the size. Security documents that arrive as attachments and never enter a calculation. A repayment test borrowed from consumer lending and pointed at a company. A price whose spread between the best and worst borrower is a fraction of the difference in expected loss.
Every one of these is a credit decision that was never taken, surfacing later as a system that cannot answer the only question that matters. A scoring engine is built to answer do we lend, yes or no. A bank has to answer something harder: how much, against what security, for what tenor, and at what price.
This is the book that answers the second question.
What is inside
Five capability layers, each with defined inputs and outputs so it can be built, tested and replaced on its own — and one rule that makes that independence real rather than aspirational.
Seven playbooks that cover every corporate segment, derived from eleven dimensions by a reduction that is proved rather than asserted. Fifteen thousand configuration cells collapse to seven, and the argument for why is given in full.
Four limit ceilings, combined by taking the minimum and never the average, because each is a real constraint and none compensates for another. A worked example runs end to end: a firm pledges eighty thousand dollars of real estate and receives twenty-nine thousand seven hundred, because that is all its cash flow can service. The rest of the collateral goes unused — which is the purpose of the formula, not a defect in it.
Two hundred and fifty-seven acceptance cases across two decision spaces, each with an expected result specific enough to compare mechanically.
Two hundred and sixty-two parameters, each with a code, a proposed value, an approving authority and a traceable source back to the chapter that defines it. Seventeen of them carry an instruction not to load them yet.
What makes it unusual
- It reports its own errors. Five findings emerged while the arithmetic was being run, and three broke a condition of the book's own coverage proof. They are set out, not smoothed away.
- The governance is redrawn for the case where the system decides. Three lines of defence rest on an assumption that fails the moment a machine makes the call.
- No jurisdiction in it. Market figures are ranges to be replaced with your own; thresholds, playbooks and parameters carry no country at all.
- Nothing is ready to use as it stands. Every value is a starting point for your own calibration, and the first page says so.
Who it is for
Chief risk officers and credit policy owners will find the thickest and least replaceable part. Board members and executives will find their decisions in about seventy pages. Architects will find a specification precise enough to build from.
Book One of two. Book Two — Building It carries the deterministic platform, three rounds of AI transformation and the full source code. Chapters are numbered continuously across both.
MOBILUCK · code247.ai
