Skip to content

Certainty & Causality Inflation Check

Inflation check — instantiates Summary-Substance Alignment Audit

Catches the summary that upgrades the substance's certainty or causality — a hedge hardened into a fact, an association reported as a cause, a subgroup generalized to everyone.

A Certainty & Causality Inflation Check watches one axis and one direction: the epistemic strength of a claim, upgraded. It compares the modality of each summary claim against the modality the body actually supports and fires only when the summary claims more — a hedge turned into an assertion ("may reduce" → "reduces"), an association turned into a cause ("linked to" → "causes"), a possibility turned into a certainty, a subgroup turned into "patients." It is deliberately blind to everything else: it does not ask whether a claim is on-topic or whether words were dropped, only whether the summary asserts a stronger epistemic status than the substance earns. That single-axis focus is what distinguishes it from a general consistency check and from an omission scan.

Example

A product team's dashboard TL;DR reads "the redesign increased retention." The underlying analysis says something narrower: retention rose in the weeks after launch, but the analysts note a simultaneous pricing change and label the result "consistent with, but not evidence of, a redesign effect." The inflation check reconstructs the body's real modality — correlational, confounded, hedged — and compares: the summary is causal and definite where the substance is associational and tentative. Its materiality rule treats a correlation-to-causation upgrade as always material, so it fires twice: once on the causal jump, once on the certainty jump. The TL;DR is rewritten to "retention rose after the redesign (alongside a price change; effect not isolated)," which is what a manager about to greenlight more redesigns actually needs to read.

How it works

  • Reconstruct the body's claims with their modality intact — its hedges, confounders, and scope — as the baseline.
  • Read each summary claim's modality on three ladders: certain↔tentative, causal↔associational, universal↔scoped.
  • Apply the materiality rule to the shift: an upgrade that would change what a reader does (correlation→causation, in-vitro→clinical, subgroup→all) is flagged; a downgrade or a lateral rewording is not.
  • Report the upgrades, each paired with the body modality it overran.

Tuning parameters

  • Materiality rule contents — which shifts are always material (correlation→causation, "some"→"all", "in mice"→unqualified); the rule is the check's engine and worth writing down explicitly.
  • Certainty-ladder granularity — how many rungs between "proven" and "possible"; finer ladders catch subtle hardening but flag more.
  • Causal-language sensitivity — how aggressively verbs like drives, causes, boosts are read as causal claims.
  • Scope-generalization sensitivity — how readily a narrowed population in the body must reappear in the summary.
  • Asymmetry — it flags upgrades and ignores faithful softening; loosening this turns it into a noisy two-way diff.

When it helps, and when it misleads

Its strength is catching the distortion that most changes behavior — the summary that makes the evidence sound more certain or more causal than it is — even when no single word was technically "dropped." The error it exists to catch is the oldest one in inference: treating correlation as causation.[1]

Its limits: it needs the body's true modality to be explicit; if the substance is itself vague, there is no baseline to grade against. Tuned too loosely, ordinary compression trips it and the alarms get ignored; tuned too tightly, it never fires. The misuse is setting the rule to wave through a favored claim's upgrade as "close enough." The discipline is to keep the rule short and anchored to decision-changing shifts, fixed before the check is run.

How it implements the components

This check fills the modality-fidelity slice of the audit:

  • material_qualifier_rule — the rule that classifies which certainty, causal, and scope upgrades are material; the check's decision engine.
  • substantive_surface_claim_set — the body's claims reconstructed with their true epistemic strength (hedges, confounders, scope), the baseline every summary claim's modality is graded against.

It judges only the direction and size of an epistemic upgrade. It does not inventory the qualifiers a summary silently dropped — that omission-first audit is Qualifier-Drop Scan — nor test whether the body simply supports the headline at all (Headline–Body Consistency Check), nor weigh a promotional author's incentive (Press-Release Claim Review).

Notes

Inflation and omission are complementary, not the same: this check fires when the summary adds strength the body lacks; the Qualifier-Drop Scan fires when the summary removes a caveat the body carried. A summary can pass one and fail the other, so the two run together — one reads the claims that are present, the other the qualifiers that are absent.

References

[1] The error the check targets is treating correlation as causation — reporting an observational association ("X was linked to Y") as though the study had shown that X causes Y. The certainty ladder generalizes the same move to "possible" reported as "proven." ↩