↩ ExploRational · A Catalogue of Divergent Rationality
explorational · a divergent move
Some facts cannot be seen from inside the reigning story — they become evidence only when a rival exists. Why does inventing an alternative, even for a theory that is working, change what the world is able to tell you?
Paul Feyerabend noticed something that ought to be impossible: certain facts are invisible from inside the reigning theory and turn into evidence only once a rival exists to make them speak. Brownian motion — the ceaseless jitter of pollen grains under a microscope — sat in plain view for decades meaning almost nothing, a curiosity of the slide. It became decisive only when the kinetic theory of matter offered a picture that predicted it and the smooth-continuum view did not. The fact never changed. The room changed. A second theory walked in, and a standing observation became testimony.
John Stuart Mill saw the mirror image in the life of a belief: even a true opinion, held with no live challenger, decays into dead dogma — still professed, no longer understood, its grounds forgotten because they never have to be defended. Together they imply a practical rule: proliferate rivals even when the incumbent appears successful. Rivals keep the evidence visible and the belief intelligible. The engine below scores theories against a bank of facts, showing which facts each theory illuminates.
the signature instrument
Each row is a fact; each column a theory; each cell that theory's honest expectation — P(fact | theory), click to edit. The one control that matters is which theories are in the room. With a single theory live, every fact's diagnostic power is exactly zero — there is nothing to compare against. Toggle a rival in and watch specific facts light up while facts both theories expect stay dark. The displayed values are live Jensen–Shannon information, in bits.
The nature of matter
6 facts · toggle rivals into the room
A belief updated by a stream of observations. The dashed line is the posterior it would hold with no rival in the room — pinned at 100%, forever, whatever the evidence. The solid line is the same belief once rivals are admitted: now it moves.
evidence is relational
A fact you can explain is not a fact that supports you. Consistent with is a one-place predicate — it interrogates no alternative — and one-place predicates carry no information. Evidence is a two-place relation: a fact is evidence for A over B, and its strength is the likelihood ratio P(fact | A) / P(fact | B). A ratio needs a denominator. With a single theory in the room there is no denominator; every observation divides by itself and equals one; the posterior cannot move. That is the arithmetic under Feyerabend's remark — the fact is not yet evidence because the term you would divide by does not exist.
Watch which facts stay dark in a full room: the ones every theory already expects. In the shipped bank, gas pressure rising with temperature and a hot body cooling read the same likelihood under all three theories, so their power is 0.00 bits however many rivals you add. They feel like confirmations and settle nothing — a fact consistent with every story cannot tell the stories apart. Scoring exactly which fact discriminates is the discipline of competing-hypotheses analysis; this move is the prior question — whether any rival is in the room at all.
what to try
Hit "one theory · all dark." Every power bar drops to zero and the total reads 0.00 bits — nothing is wrong with the facts, there is simply nothing to compare. Now toggle Kinetic back into the room. Brownian jitter and the integer-ratio law snap to 0.53 bits each; the rival created the evidence.
Feed the dogma. In the one-theory state, run the stream — the belief sits at 100% forever; the solid line lies exactly on the dashed one. Admit rivals and the same stream starts to move it: track Kinetic against a kinetic world and it climbs from a third to near-certainty. Confidence that is earned, not assumed.
Try "the mere naysayer." It puts a rival in the room whose predictions merely echo the incumbent's; the bank stays dark and the total barely leaves zero — 0.05 bits. Swap it for the structured rival and the same slot delivers 1.50 bits, over thirty times as much. Opposition is not an alternative.
proliferate even the successful
Feyerabend's point cuts against the intuition that you invent alternatives only when your theory is in trouble. The kinetic theory was not summoned because the continuum picture had collapsed; it was built alongside a picture that worked, and its arrival retroactively turned standing observations into evidence. A rival is not an assault on a successful theory. It is an organ of perception the successful theory cannot grow on its own — it lets the incumbent finally see facts that were in front of it the whole time.
Mill's complement is about what confidence is worth. A theory that has never met a live rival holds its ground by default, not by merit; its posterior sits at 100% because the denominator is empty, not because it won anything. Admit real rivals and that same theory, if it is true, earns its confidence — climbing back toward certainty by out-predicting the alternatives on evidence that can now speak. In the engine, the tracked belief starts at a third and works its way up as the stream comes in. Confirmation is not the enemy of proliferation; it is its payoff. A true theory has nothing to fear from rivals and everything to gain, because only against them is it more than dogma.
model
Each cell is a likelihood P(fact | theory) drawn from a stylized spectrum (0.1 to 0.9) that you can edit — a stipulated expectation, not a measured constant. Each fact's diagnostic power is the Jensen–Shannon divergence, in bits, summed over every pair of live theories: it is exactly the information an observation of that fact carries about which rival is right, when the rivals are held equally likely. With one theory there are no pairs, so the sum is zero — the "all dark" result is not a threshold or a fudge, it is the empty sum. Total information is that quantity added across the bank, and it can only rise as rivals enter, because every new rival contributes non-negative terms. The stream updates an ordinary Bayesian posterior over the live set; with one theory, normalising over a single hypothesis returns 1 on every step — the flat line, by construction.
The idealisations are real and worth naming. Facts are treated as independent witnesses — they are not, and double-counting is the failure the sibling move guards against. The likelihoods are stipulated, not estimated. And the engine prices a rival at zero, which is the largest fiction of all: in a scientific life, proliferating a serious alternative costs years, funding, and standing, and the incumbents rarely make room. Feyerabend's imperative is free on this page and expensive everywhere else. Read the engine as a claim about structure — that a rival is what converts a fact into evidence — never as a measurement of any real dispute.
the move ↔ the engine
| The move | The engine |
|---|---|
| a theory in the room | a story you take seriously enough to score against the facts. |
| a likelihood P(fact|H) | what that story honestly expects of each fact, committed in advance, before you look. |
| diagnostic power | what a fact can say given the rivals present — measured in bits, and exactly zero if it stands alone. |
| the dark bank | a world with one story: every fact merely "consistent," so nothing counts as evidence. |
| the lit item | the fact a rival just turned into evidence — Brownian motion, the moment kinetic theory arrived. |
| the cheap naysayer | opposition with no structure of its own; a rival that predicts nothing the room did not already predict. |
how this opening fails
risk · contrarian lock-in
The catalogue's named risk for this move. A rival defined purely against the incumbent hardens into a fixed counter-position and stops being an instrument of vision — it is still a single door, opened the other way. The engine shows the failure numerically: a naysayer whose predictions merely mirror a theory already in the room adds almost nothing — 0.05 bits against a structured rival's 1.50. Opposition feels like proliferation and is not. A rival earns its place by predicting differently, not by disagreeing loudly; a mere echo barely stirs the bank.
risk · proliferation without selection
Multiplying theories opens the evidence but closes nothing. A room crowded with live rivals and no scoring discipline is just noise with good intentions — endless deferral presented as rigour. Proliferation is the first move, not the last: the rivals it invents must then be run against the same evidence and scored, or the openness never converts into a verdict. That scoring is the work of competing-hypotheses analysis — the sibling move, and the one that closes what this one opens.
A fact becomes evidence only against an alternative — no rivals, no verdicts.
can you use it?
RECOGNITION — Your architecture is working. Which puts a rival in the room? A: waiting for the incumbent design to fail before considering alternatives. B: a colleague who opposes every decision without predictions of his own. C: a structured rival design that expects different load behavior, scored against the same facts.
C. B is the mere naysayer — an echo with a minus sign, worth almost nothing in bits. A is the intuition the move cuts against: rivals serve successful theories too.
THE NEAREST NEIGHBOR — One criterion between proliferation and competing-hypotheses analysis?
Direction. Proliferation opens — it puts rivals in the room, which is what turns facts into evidence. Competing-hypotheses analysis closes — it scores the facts once rivals exist.
PRODUCTION — Sales are flat, and "the market is saturated" explains every metric. Build a rival before looking.
Rival: pricing outgrew the mid-tier. It predicts churn concentrated at renewal dates — which saturation does not — so a standing dashboard column becomes evidence. Yours works if the rival expects something observably different from the incumbent, lighting at least one dark fact.