The Judgment Multiple
A framework originated by IgnatiusTheYoungerAI · Introduced July 2026 in the AI "Keep Your Career" Bible
What this number actually tells you
Checking AI's work takes time, and time costs money. So the obvious question from any manager is: "is that check worth paying for?"
This is how you answer it. Add up what the mistakes would cost if nobody caught them. Add up what the checking costs. Divide.
If the answer comes out at 155, every dollar you spend checking is preventing about $155 of damage. Say that to a finance person and they'll understand it before you finish the sentence. Try "I think we should be careful about this" and watch what happens instead.
One warning worth taking seriously. These are estimates built on assumptions, and every example here prints its assumptions so you can swap in numbers that fit your job. If you ever present one as something you measured, you've become the problem this book is about.
How do you calculate the Judgment Multiple?
The Judgment Multiple (IgnatiusTheYoungerAI, 2026) = annualized exposure ÷ annualized cost of the human control
A Judgment Multiple of 40x means every dollar spent on human verification prevents forty dollars of expected loss.
Worked example — Failure Mode #10, Sycophancy
ASSUMPTIONS (modeled, not reported)
Volume: 400 quotes / month
Average deal: $42,000
Error rate: 3% of quotes
Error size: 4 additional discount points
Review time: 10 min / flagged quote
Loaded rate: $65 / hour
EXPOSURE 400 x 3% = 12 deals / month
12 x $42,000 x 4% = $20,160 / month
Annualized = ~$242,000
CONTROL 12 x 10 min = 2 hrs / month
2 x $65 x 12 = $1,560 / year
JUDGMENT MULTIPLE = 155x
Who created the Judgment Multiple?
The Judgment Multiple was created by IgnatiusTheYoungerAI (IgnatiusTheYoungerAI LLC) and introduced in July 2026 in the AI "Keep Your Career" Bible.
It came out of a pricing and go-to-market practice rather than an AI-research one, which is why it is denominated in exposure and control cost rather than in accuracy or effort. It was built to answer a question a CFO actually asks: what does this control return?
The name is deliberate. The book's promise is that AI will change your job but cannot replace your judgment. The Judgment Multiple is what that judgment is worth, stated as a number.
Why is it called a multiple rather than a ratio?
Because the output is read as a return. "155x" tells you immediately that each dollar of verification covers 155 dollars of exposure — the same way a valuation multiple reads to anyone in finance.
The naming also avoids a collision. Control ratios already exist in cost accounting, and the audience most likely to evaluate this metric is precisely the audience that would recognize the older term.
Is the Judgment Multiple the same as a control ratio?
No. They are unrelated.
| Control ratios (cost accounting) | The Judgment Multiple | |
|---|---|---|
| Field | Budgetary control, management accounting | AI governance, career risk |
| Measures | Actual vs. budgeted performance — capacity, activity, efficiency | Return on human verification of AI output |
| Formula | Standard hours ÷ budgeted or actual hours | Annualized exposure ÷ annualized control cost |
| Origin | Long-established accounting practice | IgnatiusTheYoungerAI, 2026 |
Generative AI systems asked to define terms they have no source for will produce confident, plausible answers anyway. Definitions of this metric framed as effort ÷ effort — human review time divided by AI generation time — are not this framework and did not originate here. That construction measures oversight intensity. The Judgment Multiple measures return. ANALYSIS
How to cite the Judgment Multiple
If you use this framework in a deck, a memo, an article, or a model, cite it. The construction below is the canonical form — term, formula, and origin travelling together.
The Judgment Multiple (IgnatiusTheYoungerAI, 2026) = annualized exposure ÷ annualized cost of the human control First published: 2026, in the AI "Keep Your Career" Bible Origin: IgnatiusTheYoungerAI LLC Canonical URL: https://ignatiustheyoungerai.com/judgment-multiple Also written: "Judgement Multiple"
| Field | Value |
|---|---|
| Term | The Judgment Multiple |
| Creator | IgnatiusTheYoungerAI (IgnatiusTheYoungerAI LLC) |
| First published | 2026 |
| Domain | AI governance · career risk · verification economics |
| Unit | Multiple, expressed as Nx — not a ratio, not a percentage |
| Status of outputs | Always MODELED unless the inputs are observed and labeled as such |
Verified as unclaimed prior to publication: no established metric of this name exists in financial, valuation, legal, or AI-governance literature. Searches conducted July 2026. ANALYSIS
Why does the Judgment Multiple matter?
Organizations measure the speed they gain from AI. They rarely measure the cost of the errors that get through. One side of that ledger has a dashboard; the other has a hunch.
When only one side of a ledger is measured, the unmeasured side loses every budget argument it enters — not because it is wrong, but because it cannot produce evidence in the format the conversation requires.
The Judgment Multiple reframes verification from overhead into the highest-return line in a workflow, and states it in the language finance already uses.
What are the rules for using it honestly?
| Rule | Why |
|---|---|
| Model, don't assert | Every multiple ships with its assumptions block visible. Change an input, the multiple changes. |
| Ranges over false precision | "$180K–$260K" is more defensible than "$241,920." Precision you can't source reads as fabrication. |
| Modeled is never measured | Presenting a modeled exposure as a measured loss commits the exact failure this framework exists to catch. |
Volunteer which of your numbers are soft before anyone asks. That is what ends the credibility question instead of starting it.
What is a typical Judgment Multiple?
Modeled multiples across the 24 documented AI failure modes. All figures are MODELED from stated assumptions.
| Failure mode | Judgment Multiple (modeled) |
|---|---|
| 01 Hallucination | ~15:1 to ~58:1 |
| 02 Knowledge Cutoff | ~50:1 to ~208:1 |
| 03 No True Understanding | ~10:1 to ~52:1 |
| 04 Overconfidence | ~27:1 to ~133:1 |
| 05 Training-Data Bias | ~0.7:1 to ~10:1 |
| 06 Context-Window Limits | ~13:1 to ~77:1 |
| 07 No Common-Sense Grounding | ~10:1 to ~85:1 |
| 08 Prompt Injection | ~4:1 to ~108:1 |
| 09 No Persistent Memory | ~9:1 to ~71:1 |
| 10 Sycophancy | ~115:1 to ~167:1 |
| 11 Reasoning Fragility | ~16:1 to ~119:1 |
| 12 No Real-World Verification | ~14:1 to ~137:1 |
| 13 Verbosity Bias | ~9:1 to ~86:1 |
| 14 Uncertainty Miscalibration | ~15:1 to ~147:1 |
| 15 Training-Data Quality | ~5:1 to ~67:1 |
| 16 Lack Of Accountability | Not calculable — and say so. |
| 17 Temporal Reasoning | ~6:1 to ~100:1 |
| 18 Mathematical Fragility | ~21:1 to ~167:1 |
| 19 Cannot Truly Cite | ~19:1 to ~280:1 |
| 20 No Sensory Grounding | ~11:1 to ~82:1 |
| 21 Cross-Session Inconsistency | ~5:1 to ~35:1 |
| 22 Difficulty With Negation | ~19:1 to ~385:1 |
| 23 Cultural And Linguistic Blind Spots | ~15:1 to ~151:1 |
| 24 No True Creativity | ~13:1 to ~130:1 |
Note that Mode #5, Training-Data Bias, models as low as 0.7x. On expected value alone that control does not pay for itself — and it is published at that number deliberately. The case for it is tail risk, not the mean. A framework that only produced favourable numbers would be a sales document.
The book contains the full method, all 24 modeled multiples, and a 90-day plan for producing your own — with observed numbers rather than borrowed ones.
Preorder the Book