Hypothesis Screening

Reduce a candidate list of hypotheses to a small set of survivors — genuine rivals worth testing — through joint, criterion-by-criterion elimination rather than a weighted score or a popularity vote.

The problem

Researchers facing time or resource constraints often have more candidate hypotheses than they can test. The temptation is to keep whichever hypotheses are individually most plausible. This produces a shortlist that may contain near-duplicates (hypotheses that would predict the same thing) or hypotheses that cannot actually be evaluated with available or attainable evidence. The task here is not to rank hypotheses by average appeal, but to eliminate any candidate that fails decisively on any one of several independent criteria, leaving only hypotheses that are testable, minimal, plausible, distinct from one another, and capable of predicting something beyond the pattern that motivated them.

Step 0 — Elicit and normalize the candidate list

Ask for, or extract, the full list of candidate hypotheses. For each one, restate it as a single sentence that names the variables or constructs involved. If a candidate is vague, compound, or conflates two mechanisms, split it or ask the user to disambiguate before screening — a hypothesis cannot be screened against Criterion 6 (specificity) if it isn't yet stated specifically.

If more than ~8 candidates are supplied, screen all of them; do not pre-filter informally before applying the criteria.

Step 1 — Apply the five criteria jointly

Evaluate every candidate against all six criteria below. These are applied jointly, not sequentially and not as a weighted average. Strength on one criterion cannot compensate for clear weakness on another.

1. Testability — Is there any observable evidence, even indirect, that bears on this hypothesis? A hypothesis with no conceivable observable implication is cut regardless of its other merits.
2. Parsimony — Does it explain the pattern without invoking extra unobserved machinery (unobserved types, ad hoc moderators, multiple undetected shocks)? Penalize hypotheses that only work by adding unfalsifiable auxiliary assumptions.
3. Causal-density fit — In a domain where many causes plausibly produce one outcome, is this a serious candidate mechanism, or is it a strawman included only to be knocked down? Judge plausibility against the actual causal structure of the domain, not against how easy it is to state.
4. Discriminant validity — Is the candidate meaningfully distinct from the other survivors, or does it collapse into the same mechanism as one you're already keeping? This is evaluated relative to the emerging survivor set, not in isolation (see Step 2).
5. Independent predictive reach (anti-HARKing check) — Does the hypothesis imply anything beyond the pattern that triggered it? Fit to the originating data alone does not count as support; the hypothesis must generate a prediction about something not yet observed.
6. Specificity — Does the hypothesis statement clearly name the variables of interest (what causes what, for whom, under what conditions)? Vague or compound statements are cut or sent back for restatement.

Step 2 — Eliminate sequentially by discovery order, not by final score

• Cut a candidate the moment it is clearly weak on any criterion. Record which criterion it failed. Do not average across criteria and do not let a high score on Criteria 1‚Äì3 rescue a candidate that fails Criterion 5 or 6.
• Evaluate Criterion 4 against the survivor set as it stands, not against the full original list. This means order matters: if a candidate looked like a keeper before a stronger, later candidate turned out to explain the same variation, the earlier one is displaced, not kept alongside it.
• If two strong candidates are tied on discriminant grounds ‚Äî they would predict the same things ‚Äî keep only the one stronger on testability or causal-density fit and drop the other. Do not keep both merely because both are individually reasonable.
• Show the elimination reasoning for every cut candidate; do not silently drop hypotheses.

Step 3 — Confirm the survivor set is a genuine rivalry

Stop at exactly three survivors. Before finalizing, check that each survivor is a genuine rival to the other two: there must be at least one Criterion 5 (predictive-reach) test on which the three would diverge from one another. If any pair of survivors would make identical predictions on every attainable test, they are not yet distinct — return to Step 2 and resolve the collision (displace one, or find the test that separates them) before finalizing.

Required output

Return the following, in order:

1. Normalized candidate list — each hypothesis restated as a single, specific sentence naming its variables.
2. Elimination log — for every cut candidate: the criterion it failed and a one-line justification. For displaced candidates, note what later candidate displaced them and why.
3. Survivor set (exactly 3) — each survivor stated precisely, with a brief note on why it passes all six criteria.
4. Divergence check — for each pair of survivors, name at least one attainable test/observation where their predictions would differ.
5. Priority order — if the three survivors are not equally cheap or equally informative to test, rank them for testing order and say why.

Quality checks

• Never keep more or fewer than three survivors without flagging to the user that the candidate set doesn't naturally support three genuine rivals (e.g., fewer than three testable, non-collapsing candidates exist).
• Never justify keeping a candidate by its performance on only some criteria; state explicitly that all six were checked.
• Never treat fit to the triggering pattern as evidence in the Criterion 5 check ‚Äî that criterion requires a prediction beyond the original data.
• Never let two survivors stand if they are observationally equivalent on every attainable test; the divergence check in Step 3 is mandatory, not optional.
• Do not silently reorder or omit candidates from the elimination log.