Why it matters

An ACH matrix — short for Analysis of Competing Hypotheses — is a grid that pits several explanations against all the evidence at once. The competing explanations run across the top as columns; the pieces of evidence run down the side as rows; and every cell records one small judgment: is this piece of evidence consistent with that explanation, inconsistent with it, or simply not relevant either way. The point of laying it out this way is the opposite of what most people do with an argument. You don’t pick the explanation the evidence supports — you pick the one the evidence fails to contradict. The matrix is built to make a team reason by elimination, because the explanation you can’t knock down is a safer bet than the one you happened to like first.

For example: an incident team is trying to explain why a service’s error rate tripled overnight. The on-call engineer is sure it was the deploy that went out that evening, and the room is ready to roll it back. Laid out as an ACH matrix — columns for “the deploy,” “a dependency outage,” “a traffic spike,” “a config change” — the same evidence forces a mark in every cell, and one row settles it: the error rate had already started climbing two hours before the deploy shipped. That single fact is inconsistent with the deploy theory and inconsistent with the traffic-spike theory, but perfectly consistent with the dependency outage. The deploy was the obvious suspect. The timing row was the piece of evidence nobody had lined up against all four stories at once.

  • What it shows. Every competing explanation scored against every piece of evidence, so you can see at a glance which explanations the evidence rules out — not just which one it flatters.
  • When to reach for it. Several live explanations compete, the evidence is mixed enough that no single one obviously wins, and the cost of believing the wrong one is high enough to be worth the discipline.
  • How to read it. Don’t count the ticks — count the crosses. Scan down each explanation’s column and tally its inconsistencies; the column with the fewest is the one left standing. Then look across the rows for the evidence that splits the explanations apart.
  • What you’d miss without it. The evidence that quietly contradicts your favored story. In prose, an inconvenient fact gets glossed; in a grid, it sits in a cell with a cross in it that you have to look at.
  • Where it misleads. A clean column can feel like proof when it’s only “not yet falsified” — the matrix narrows the field, it doesn’t certify a winner. And the marks are judgment calls: a wishful analyst can soften a damning “inconsistent” into a comfortable “not relevant” and quietly rig the count.

How to read it

Picture a spreadsheet. Across the top, each column is a hypothesis — one candidate explanation, written out as a full sentence so it can actually be tested (“the price increase changed who was buying,” not just “pricing”). Down the left, each row is a piece of evidence — one fact, observation, or report, also stated precisely enough to argue with. Where a column meets a row, the cell holds a single mark: the evidence is Consistent with that hypothesis, Inconsistent with it, or Not applicable — it doesn’t bear on that hypothesis one way or the other. Tools often render these as colors too: green for consistent, red for inconsistent, grey for not-applicable, so the pattern jumps out before you read a word.

Now the move that makes this diagram different from every ordinary comparison: you score by counting the inconsistencies, not the consistencies. Walk down a single hypothesis’s column and tally its red cells. The hypothesis with the fewest inconsistencies is the one that survives. This feels backwards the first time, and it’s the whole idea. Lots of facts are consistent with lots of explanations — “the error rate was high” fits almost any theory of why — so consistent evidence rarely tells the explanations apart. But a single inconsistency is decisive: if the evidence genuinely contradicts a hypothesis, that hypothesis is in trouble no matter how much else fits. So the matrix rewards the explanation that steps on the fewest facts. This is disconfirmation, not confirmation — the same instinct as a scientist trying to break a theory rather than cheer for it.

That reframing changes what counts as a valuable piece of evidence, and the rows make it visible. A row that reads C across every column — consistent with all of them — is doing nothing; it can’t move you toward any hypothesis, so it’s filler dressed up as support. The rows that earn their place are the diagnostic ones: a row that’s inconsistent with three hypotheses and consistent with the fourth has, in one stroke, eliminated three and pointed at one. So read the matrix twice. Down the columns to find the surviving explanation; across the rows to find the handful of facts that actually did the discriminating — those are the ones worth double-checking, because the verdict is resting on them.

When to use it

The ACH matrix is the visible artifact of the HYPOTHESIS / EVALUATION family — the diagrams whose job is to weigh competing explanations against evidence rather than to map causes or rank options. It is specifically the which-explanation-is-true tool, and that boundary is how you pick it over its neighbors. A decision matrix (sometimes a weighted scoring grid) looks superficially identical — a grid with things across the top and things down the side — but it answers a different question: it scores options against criteria to find the best choice (“which vendor should we pick,” weighing price, support, and fit). ACH scores hypotheses against evidence to find the most defensible belief (“what actually happened”). Choice versus truth — that’s the line. And a plain pro/con list handles the simplest version of the same instinct: a single option weighed for and against, with no second hypothesis to compete with it and no shared evidence to discriminate between rivals.

Reach for an ACH matrix when there are at least three explanations still genuinely in play, the evidence is mixed enough that the easy answer might be wrong, and being wrong would be expensive — intelligence assessments, incident root-causing, fraud examination, medical differential diagnosis, investigative reporting, historical reconstruction. Skip it when the evidence overwhelmingly points one way (you don’t need a grid to tell you the obvious), when the question is really binary (a single pro/con or a two-row truth table is enough), or when you’re choosing rather than diagnosing — wanting the best option, not the true explanation, is a decision matrix’s job. The matrix is the tool for the moment when you suspect your first guess is the comfortable one rather than the correct one.

How Ora builds it

Ora produces an ACH matrix from a semantic spec — a structured description rather than a hand-drawn picture. The spec carries a list of hypotheses (each a full-sentence explanation), a list of evidence items (each a fact with its source and a note on how credible and how relevant it is), and the consistency rating for every hypothesis-evidence pair: consistent, inconsistent, or not-applicable. On top of that it records two derived things — the diagnosticity of each evidence row (how sharply it discriminates between hypotheses; a row that varies a lot across the columns is doing real work, a row that’s uniform is not) and the per-hypothesis score, the count of inconsistencies that decides which explanation survives. Some specs also weight the marks, treating a flat-out contradiction as heavier than a mild one, so the score is a weighted tally rather than a raw count.

That spec is then rendered as a grid — a table for inline reading, or a heatmap-style colored matrix where each cell’s shade encodes its mark and the inconsistency totals are summed at the foot of each column. (Because the meaning can’t live in color alone, the symbols carry it independently and the rendering ships with alt-text that names the column tallies and calls out the most diagnostic rows, so the grid is legible to a screen reader and not only to the eye.) The diagram is the face of Ora’s competing-hypotheses mode: when you ask it to weigh several explanations against your evidence and tell you which holds up, that mode runs the cell-by-cell consistency sweep, scores by elimination, and emits this artifact to show its working.

The technique is the work of Richards J. Heuer Jr., a veteran CIA analyst who set it out in Psychology of Intelligence Analysis (1999), written for the agency’s Center for the Study of Intelligence and now in the public domain. It grew out of the discipline’s hard-won lesson that the dangerous error isn’t ignorance but confirmation bias — latching onto the first plausible story and reading every new fact as support — and it carries the fingerprints of Karl Popper’s idea that a theory earns its keep by surviving attempts to falsify it. The method has since spread well past the intelligence world into the broader practitioner literature on structured analytic techniques, where Heuer and Randolph Pherson’s Structured Analytic Techniques for Intelligence Analysis (2010) is the standard reference, and into CIA-tradecraft-style red-team training used in business strategy, fraud examination, and competitive intelligence.

  • Heatmap — the same colored-grid rendering ACH borrows for its cell coding; a heatmap shows magnitude across two dimensions, where ACH overloads the grid with a three-way consistency judgment.
  • Quadrant Matrix — the other classification grid in the family, sorting items into four boxes on two axes; ACH trades the two clean axes for an open set of evidence rows scored against an open set of hypotheses.
  • Decision Tree — the sibling for choosing under uncertainty (sequential options and probabilistic outcomes) rather than diagnosing which explanation is true.
  • Analysis of Competing Hypotheses (mode) — the analytical operation that runs the cell-by-cell consistency sweep and scores by elimination; this matrix is how that mode shows its work.