Reasoning Strategies, Part 3: The Scientific Family
Part 1 covered the epistemic family (what follows from what); Part 2, the problem-solving family (how to reach a goal defensibly). This installment releases the scientific family in the Reasoning Library: two strategies that govern how belief is earned and revised. The hypothetico-deductive method makes an explanation pay for credibility by surviving a genuine attempt to break it. Bayesian updating makes every piece of evidence pay for its influence in diagnosticity, with the prior on the record and the belief trajectory auditable. Both target the same enemy from different sides: confirmation absorbed into process.
What this family is for
Investigations and monitoring are the two places regulated institutions manufacture belief at scale. An analyst closes an alert believing the activity is explained; a periodic review reaffirms a rating believing nothing has changed. The scientific strategies are the discipline of that manufacturing: they specify when a belief has earned its confidence, and — just as important — what would revoke it.
The family’s shared enemy is confirmation, and it has two industrial forms. The first is the theory that explains everything — an investigation narrative so flexible that every new fact confirms it, which means no fact tests it. The second is the ratchet — the risk rating that only ever rises, because each vivid detail is counted as weight and nothing is charged with moving belief the other way. Both feel like diligence from the inside. Both are the specific failure the family’s gates exist to catch — and both are failures an unguided language model reproduces enthusiastically, because completing the running story is exactly what next-token generation does.
Hypothetico-deductive method — predict, then try to break it
The working loop of experimental science, run as a graph: state the hypothesis, deduce observations that must hold if it is true — and that would look different if it were false — then retrieve the evidence in both directions. A dedicated critique node asks the only question that matters: did anything break it? The gate admits three verdicts and only three: refuted, corroborated (survival of the named tests — never “proven,” and the trace says so), or untested, which routes to escalation instead of masquerading as a finding.
The regulated fit is investigations: structuring theories, fraud typologies, root-cause narratives after an incident. Each is a hypothesis that predicts things beyond the alert that raised it, and the graph’s output is what a defensible SAR narrative — or a defensible decision not to file — actually requires: the rival explanations, the discriminating evidence pulled, and what it showed.
Deep dive: how the hypothetico-deductive graph works, step by step, with three grounded use cases →
Bayesian updating — belief in proportion to evidence
Belief revision with bookkeeping: the prior is recorded with its source before any case detail is admitted; each evidence item enters once, deduplicated at the door; an evaluation node prices each item by diagnosticity — how much more likely the observation is under one hypothesis than the other — and belief moves proportionally, in either direction. The calibration gate audits the run for the two classic corruptions: the same fact counted under three phrasings, and the trajectory that only ever pointed one way.
The regulated fit is monitoring: KYC refresh, enhanced due diligence, alert scoring, early-warning credit review. “Why is this customer high-risk?” stops being a narrative and becomes a decomposition — this prior, these items, these weights — which is simultaneously the answer for the examiner and the challenge surface for model validation. Its honest limit is stated in the design: priors and weights are judgments. The graph does not hide that; it records them, which is what makes them challengeable.
Deep dive: how the Bayesian graph works, step by step, with three grounded use cases →
Two strategies, one discipline
The pair compose naturally: abduction (Part 1) proposes candidate explanations, the hypothetico-deductive graph tests the leading one, and Bayesian updating carries the surviving belief forward through time as evidence accumulates. What the family adds to the inventory is a governance property none of the earlier strategies supply on their own: revisability with receipts. Every belief in the system carries what would change it — the prediction that would refute it, the evidence that would move it — so “we were wrong” becomes a state transition with a trace, not a crisis with a narrative.
With this release, four of the seven families are fully published — thirteen graphs, each with its walkthrough, gates, and worked use cases. Next in the series: the design and linguistic family — Double Diamond, lateral thinking, schema instantiation, and pragmatic inference.
The strategy library: series index
The full registry — thirteen live graphs across four families, with three families to come — lives in the Reasoning Library, with every graph one click from its deep dive. The two released in this installment:
A belief that names what would change it is governance. A belief that doesn’t is a commitment — and the file can no longer tell the difference. The scientific family exists to keep that distinction mechanical.