Falsifiability Checker

Falsifiability Checker — The Popperian Translator

A thin coordinator that translates a problem statement into the claim-shaped input that demarcation-judge (philosophy/philosophy-of-science) evaluates, then translates the output back into binding-vow's grader-axis terms.

The actual falsifiability assessment is done by demarcation-judge. This skill's value is the translation — problem statements rarely arrive shaped as claims, and feeding them to demarcation-judge directly produces nonsense. The translation step is small but load-bearing.

Process

Step 1 — Extract implicit claims

Most problem statements embed claims rather than make them explicit. Extract:

• Predictive claims ("if we do X, Y will happen") — directly falsifiable
• Causal claims ("X is causing Y") — falsifiable if mechanism is specified
• Existence claims ("Y is happening to N% of users") — falsifiable via measurement
• Comparative claims ("X is worse than Z") — falsifiable via head-to-head data

If the statement is purely interrogative ("how do I do X?") with no embedded claim, return verdict N/A — interrogative and skip to output.

Step 2 — Reformulate as claims and pass to demarcation-judge

For each extracted claim, pass to demarcation-judge with:

• The claim text (in declarative form)
• The domain context (clinical, software, business, etc.)
• The implicit prediction the claim makes

Receive back the falsifiability verdict per claim.

Step 3 — Aggregate and translate

Translate demarcation-judge's output back into binding-vow's terms:

| demarcation-judge verdict | binding-vow translation | Falsifiability axis score | |---|---|---| | Falsifiable, makes risky predictions | Empirically grounded | 5 | | Falsifiable but cautious; predictions are weak | Acceptable | 4 | | Falsifiable in principle but practically untestable | Borderline | 3 | | Probabilistic / hedged; falsification requires base-rate context | Weak | 2 | | Unfalsifiable — tautology, ad hoc, or values-shaped | Fail | 1 |

If the statement contains multiple claims with mixed verdicts, the axis score is the minimum across claims (the weakest link drives the score).

Output Format

FALSIFIABILITY — [first 60 chars of statement...]
─────────────────────────────────────────────
Implicit claims extracted: [list]
Per-claim verdicts: [from demarcation-judge]
Aggregate score: [1-5]
Reformulation hint (if score < 3): [what would make this falsifiable]

Output Contract for six-eyes

Returns the falsifiability axis score (1-5) and reformulation hint for statement-grader. If aggregate score < 3, the re-state loop should reformulate the statement to make at least one claim falsifiable.

Scope Boundaries

• falsifiability-checker handles: translating problem-statement format ↔ claim format, aggregating multi-claim verdicts.
• falsifiability-checker does NOT: perform the falsifiability assessment itself — that's demarcation-judge's job. If demarcation-judge is unavailable, return "deferred" rather than guessing.

Failure Modes

| Failure | Response | |---|---| | No claims extractable from statement | Return N/A — interrogative or pure description; mark falsifiability axis as not-applicable | | demarcation-judge unavailable | Return "deferred — call demarcation-judge directly"; do not fabricate the verdict | | Statement contains a values claim ("we should X") | Pass to demarcation-judge anyway; expect verdict "unfalsifiable as values-shaped" and translate accordingly |

Connections

• statement-grader (binding-vow) — feeds the falsifiability axis
• demarcation-judge (philosophy/philosophy-of-science) — primary cross-domain call
• argument-analyst (philosophy/logic) — adjacent skill for evaluating the argument that supports the claim

Sources

• Popper, K. R. (1959). The Logic of Scientific Discovery. (Falsifiability as the demarcation criterion.)
• See [[kahneman-framing]] for the cognitive failure modes (especially attribute substitution) that produce unfalsifiable problem statements.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: npbuilds
• Source: npbuilds/skill-library
• License: MIT
• Homepage: https://skill-library-prod.web.app/

Install and usage instructions live in the source repository linked above.