Skip to content

Experience Prediction Matrix

Rival-prediction matrix tool — instantiates Sense-Experience Reduction Protocol

Tabulates the experiences each rival claim predicts, exposing the conditions under which their predictions diverge.

Experience Prediction Matrix is a grid: rival claims down one axis, candidate observations across the other, and in each cell the experience that claim predicts under that observation. Its defining job — distinct from every sibling — is comparison across competitors. A single claim can look well-supported until you lay a rival beside it and find they predict the very same experiences; the matrix's whole point is to find the cells where the rivals disagree, because those are the only observations that can tell them apart. It turns "which of these is true?" into the sharper "which observation would separate them?" — locating the discriminating experience rather than piling up evidence each rival predicts equally well.

Example

A patient presents with jaundice, fatigue, and dark urine. Three rival explanations are on the table: viral hepatitis, a gallstone blocking the bile duct, and haemolysis (red cells breaking down). An Experience Prediction Matrix lays them out against candidate observations. Ultrasound of the duct: dilated for gallstone, normal for the other two — a discriminating cell. Bilirubin fractionation: mostly conjugated for obstruction, mixed for hepatitis, mostly unconjugated for haemolysis — discriminating across all three. Liver enzymes: sharply raised for hepatitis, modest for obstruction — partly discriminating. Fever: possible under hepatitis, uncommon under haemolysis — weakly discriminating.

Read down the columns, the matrix shows that bilirubin fractionation separates all three rivals at once, while fever barely separates any. The clinician now knows which single observation buys the most discrimination — and orders that test first, rather than the one that merely confirms whichever diagnosis was already suspected.

How it works

  • Enumerate the rivals, not just the favourite. List the competing claims explicitly, including the dull "nothing unusual" alternative, so the comparison isn't rigged toward one.
  • Fill each cell with a predicted experience. For every rival × observation pair, state what that claim says the observer would encounter.
  • Read for divergence. Find the observations whose cells differ most across rivals; those carry the discriminating power. Rows that read identically across the board are worthless for choosing.
  • Rank observations by separating power. Order candidate observations by how many rivals they split apart, so the most diagnostic one can be run first.

Tuning parameters

  • Rival set completeness — how many competing claims are entered. A wider set is fairer and catches surprises but grows the grid and dilutes focus.
  • Observation candidate pool — how many possible observations are considered. More candidates find sharper discriminators but cost effort to predict for every rival.
  • Discrimination metric — how "separating power" is scored (splits any two rivals vs. splits all of them). The choice steers which observation the matrix recommends.
  • Prediction confidence — whether cells hold point predictions or hedged ranges. Ranges are honest but blur the divergences the matrix exists to expose.
  • Cost weighting — whether an observation's separating power is traded against how hard it is to obtain. Weighting favours the cheap-but-discriminating cell over the ideal-but-infeasible one.

When it helps, and when it misleads

Its strength is that it kills confirmation-seeking at the root: it steers effort toward the observation that separates hypotheses rather than the one that flatters the favourite, and it makes visible the uncomfortable case where two rivals predict identically. This is the practical logic of the crucial experiment — a single observation on which competing accounts make opposite predictions.[1]

Its failure modes are twofold. First, the matrix only compares the rivals listed; the true explanation may be a column that was never entered, so a claim can "win" by default against a weak field. Second, when rivals predict the same experiences across every affordable observation, they are empirically underdetermined — the matrix correctly reports that no available observation can decide between them, a result easy to mistake for one rival being confirmed. It is also run backwards when the observation set is pruned to leave only cells the favoured claim wins. The discipline is to include a genuine field of rivals plus the null, and to treat all-identical rows as a verdict about the claims' distinguishability, not about their truth.

How it implements the components

  • rival_claim_matrix — its structural core: the explicit array of competing claims against candidate observations.
  • counterfactual_experience_conditionals — populates each cell with the experience a given rival predicts under a given observation.
  • contrastive_disconfirmation_set — the divergent cells constitute mutual disconfirmers: an observation that matches one rival counts against the others.

It does not build the disconfirmation set for a single claim in isolation (Disconfirming Condition Probe), do the first-pass rewrite of one claim's conditionals (Counterfactual Conditional Translation), or design the procedure that runs the discriminating observation (Observation Protocol Design).

References

[1] An experimentum crucis — an observation on which two competing hypotheses make opposed predictions, so its outcome tells them apart (a notion running from Bacon through Newton). The matrix is a device for finding such observations, subject to the caveat that rivals predicting identically admit no crucial experiment at all. ↩