Skip to content
Final StateRange Compression: Five Buckets, a Thousandfold Gap
VOL. I  ·  NODE 109▢  ATLAS

ONE CELL, TWO WORLDS

Range Compression: Five Buckets, a Thousandfold Gap

Range compression is what happens when a risk matrix forces continuous probabilities and losses into a few ordinal buckets.

THE BINS THROW AWAY THE MAGNITUDE

Broad buckets can hide orders of magnitude

Ordinal probability and loss bins hiding orders-of-magnitude differences.Under the illustrated boundaries, the rare row groups 10^-3 with 10^-6 and the catastrophic column groups EUR100M with EUR10B.FREQUENCY / YRIMPACT / LOSS10^010^-110^-210^-310^-410^-510^-6EUR 10BEUR 1BEUR 100MEUR 10MEUR 1MRARE BINx1000TOP BINx100RARE + CATASTROPHICILLUSTRATIVE BINSMAGNITUDE LOST INSIDE EACH BIN
  • A broad 'rare' row can span 10⁻³ to 10⁻⁶
  • A broad 'catastrophic' column can span €100M to €10B
  • The boundaries determine what magnitude is discarded

This is the hidden cost of the frequency-impact matrix's ordinal scales: a decision about the rare and severe made on a ruler with no fine marks.

THE MAN WHO DID THE MATHS

The man who wrote it down as theorems

Source panel for Tony Cox's 2008 Risk Analysis critique of risk matrices.The supporting figure anchors the theorem beat: Cox tested consistency, betweenness, and colouring assumptions users often bring to matrices.RISK ANALYSISVOL. 28 / 2008WHAT'S WRONGWITH RISKMATRICES?TONY COXDECISION SCIENTISTCONSISTENCYBETWEENNESSCOLOURINGRISK ANALYSIS 28(2) / 2008
Cox, Risk Analysis 28(2), 2008, pp. 497–512; DOI 10.1111/j.1539-6924.2008.01030.x.
  • Louis Anthony (Tony) Cox, a decision scientist
  • Risk Analysis, 2008
  • Conditions: weak consistency, betweenness, consistent colouring

Cox's critique matters because it tests common frequency-impact matrix designs against explicit mathematical properties rather than relying on intuition.

WHAT'S WRONG WITH RISK MATRICES?

What's Wrong with Risk Matrices?

Louis Anthony (Tony) Cox, Risk Analysis, 2008

The title was a question. Cox showed that common matrices can violate desirable ordering properties and, under stated conditions, assign higher ratings to quantitatively smaller risks.

A THOUSANDFOLD GAP, FOLDED FLAT

A thousandfold gap, folded flat

Accordioned matrix buckets showing a thousandfold frequency gap and hundredfold loss gap folded flat.The data exhibit opens the row and column buckets to reveal the hidden magnitude gaps that a five-by-five score can erase.CATASTROPHICRARERARE BIN10^-310^-6x1000CATASTROPHICEUR 100MEUR 10Bx100EXPECTED LOSSEUR 100KEUR 10KSCORE 5SCORE 5DIFFERENT VALUES / SAME ORDINAL SCORE
illustrative bins and values; expected-loss arithmetic shown explicitly. Cox (2008) supplies the general critique.

Where frequency and severity pull against each other, ordinal scoring can tie or invert a quantitative ordering, depending on the bins and decision criterion.

  • The hypothetical rare row hides a 1000× probability spread
  • The hypothetical catastrophic column hides a 100× consequence spread
  • Expected loss is €100k versus €10k, yet the ordinal score ties

The coin comparison belongs to a condition: coarse ordinal bins, rank scoring, and risks where rarity rises with severity.

That is Cox's warning, not a slogan against every grid: under those tail conditions, the matrix can invert priorities and hide prevention's value at the edge of the grid. This is why the cell is not a decision.

THE WORDS DON'T MEAN THE SAME THING

The words don't mean the same thing

Verbal probability labels spreading across different numeric interpretations.The figure separates documented variation in verbal-probability interpretation from a schematic reminder to measure inter-assessor scatter.LIKELY0%100%TENS OF POINTS APARTSCHEMATIC 5 x 5SCHEMATIC / NOT STUDY DATAMEASURE ASSESSOR AGREEMENT
source-based: Budescu, Broomell & Por (2009); Ball & Watt (2013), DOI 10.1111/risa.12057. Lower grid is schematic.

Ordinal labels are weak supports for arithmetic (Hubbard & Evans, 2010) — the same discretising flaw the matrix inherits from its buckets.

  • Readers interpret verbal probabilities differently (Budescu et al., 2009)
  • Risk-matrix ratings can vary widely across assessors (Ball & Watt, 2013)
  • Numeric definitions and quantitative analysis help; they do not remove model uncertainty

ANY BIN IS A DECISION

Any bin embeds a modeling choice

Risk tail and menu choices both compressed by bins drawn upstream.The comparison links risk buckets to menu framing: whoever draws the bins shapes what differences can still be noticed.CONTINUOUS FIELDBINTHREE MENU ROWSOPTION AOPTION BOPTION CONE TAIL CELLBOUNDARIES ARE MODEL CHOICESMAKE THE DISCARDS VISIBLE
  • A tail summarized in one cell
  • A feasible set summarized in three menu rows
  • Boundary choices shape which differences remain visible

The same compression that can flatten a tail can narrow a menu into three steers. A shortlist may look neutral because it is a familiar format; audit its boundaries and discards through the illusion of choice and agenda control.

LET THE LINE RUN CONTINUOUS

For tail decisions, show the quantities behind the bin.

  • Use quantitative probability and consequence ranges where support permits
  • Keep the grid only for a first rough sort
  • Show assumptions, uncertainty, and sensitivity—not just a smooth line

What is new here is the binning math. If one cell hides order-of-magnitude differences in probability or loss, it is insufficient on its own. Keep the matrix for rough triage; for consequential tail cases, expose quantities, assumptions, and sensitivity. Back to the cell that is not a decision.

Read the transcript

01 · ONE CELL, TWO WORLDS

Take the matrix apart at the level of resolution. Range compression occurs when quantitative probabilities and consequences are forced into a few ordinal buckets. Consider a hypothetical matrix whose broad bins place one-in-a-thousand with a hundred-million-euro loss and one-in-a-million with a ten-billion-euro loss in the same cell. Their expected losses are one hundred thousand euros and ten thousand euros. The shared ordinal score hides rather than preserves that ranking. The example depends on the stated bins; it is an illustration, not Cox's dataset.

02 · THE BINS THROW AWAY THE MAGNITUDE

Under the illustrated boundaries, a broad rare row spans annual probabilities from one in a thousand to one in a million, a factor of a thousand. A broad catastrophic column spans losses from one hundred million euros to ten billion, a factor of one hundred. Other matrices may draw different boundaries. The general mechanism is the same: every bin preserves differences between categories and discards differences within them. The operating question is whether the discarded magnitude matters to this decision.

03 · THE MAN WHO DID THE MATHS

In 2008, decision scientist Louis Anthony Cox published What's Wrong with Risk Matrices in Risk Analysis. He formalized desirable properties including weak consistency, betweenness, and consistent coloring, then showed that common matrices can violate them. His critique is conditional and mathematical, not a claim that every grid always fails. In particular, coarse categories and ordinal scoring can leave many quantitative risk pairs indistinguishable and can misorder risks under stated conditions.

04 · WHAT'S WRONG WITH RISK MATRICES?

The title was a question: What's Wrong with Risk Matrices? Cox's answer was not simply the word tail. Common designs can lose resolution, violate ordering properties, and under stated conditions assign a higher qualitative rating to a quantitatively smaller risk. That is enough to reject the cell as a final decision rule while preserving a possible role for rough triage.

05 · A THOUSANDFOLD GAP, FOLDED FLAT

Watch the hypothetical compression. The chosen rare row spans a thousandfold probability difference. The chosen catastrophic column spans a hundredfold consequence difference. The first pair has expected loss of one hundred thousand euros; the second, ten thousand. Yet multiplying their shared ordinal row and column ranks produces the same score. The exhibit does not prove expected loss is the only decision criterion. It proves that this score has discarded inputs that a decision-maker may need.

06 · WORSE THAN A COIN

Cox's sharper result is conditional. It applies when a matrix uses coarse ordinal bins, users treat bin ranks as numbers, and frequency and severity move against each other, as they can in tail cases. Under those conditions, the matrix can invert the ordering of risks; the coin comparison belongs there, not everywhere. The takeaway is not that every grid is useless. It is that a cell hiding large differences in likelihood and loss may be unreliable as a final ranking rule.

07 · Advertisement · 11x AI Growth Workers

Pipeline is not one task. It is a chain of small handoffs: find the buyer, research the account, respond fast, qualify cleanly, follow up on time. 11x gives sales, marketing, and RevOps teams digital workers for that background motion. The system keeps work moving, while people focus on the conversations that deserve them.

08 · THE WORDS DON'T MEAN THE SAME THING

A second problem is interpretation. Budescu, Broomell, and Por studied how readers interpreted the IPCC's verbal probability language and found substantial variation, supporting combined verbal and numerical communication. Ball and Watt found that different assessors could assign widely scattered matrix ratings to the same hazards, with scatter remaining after reflection. The operating response is to define terms numerically, measure agreement, and expose uncertainty.

09 · ANY BIN IS A DECISION

Binning is a modeling choice. Boundaries determine which differences remain visible, whether in a matrix or a shortlist. That does not make every boundary a hidden verdict; categorization can be useful and necessary. The governance question is whether the boundaries, excluded alternatives, and sensitivity of the result are visible to the person deciding.

10 · LET THE LINE RUN CONTINUOUS

For a tail decision, inspect the quantities behind the cell. If cases differ by orders of magnitude, the cell is a triage label, not a sufficient decision object. Use probability and consequence ranges, expected-loss calculations where appropriate, exceedance curves, and sensitivity analysis. A smooth curve is not truth: it still depends on data, assumptions, and uncertainty. The remedy is not decoration. It is preserving the quantitative distinctions and showing how robust the decision is to what remains unknown.

01 / 10 · ONE CELL, TWO WORLDS0:00 / 5:60