Install Scientific Falsification Protocol in Any AI Agent

Falsify is a Markdown reasoning framework for organizing hypotheses and external evidence. It includes reference material and a thinking-ledger template. It does not guarantee accuracy, calibration, or compatibility with every assistant. Its named concepts are prompts for review rather than a validated combined scientific protocol.

Its central rule is:

NO VERDICT WITHOUT A FALSIFIABLE HYPOTHESIS.

For empirical claims, identify observations that could change the conclusion. Definitions, mathematics, values, and underdetermined evidence need an appropriate validation method rather than a forced falsification exercise. Internal reflection alone adds no new evidence.

Route the request before applying the detailed stages. Simple handles trivial or one-lookup questions directly. Nudge gives a low-stakes ballpark with its main assumption and at most a few targeted questions. Question asks for genuinely blocking inputs. Depth addresses consequential recommendations, uncertain architecture, correctness gates, and explanations of failures.

Incident mode follows the authorized runbook with bounded reversible mitigation, a rollback, predicted effects, and immediate re-observation. It prioritizes stabilization instead of demanding certainty first. Once stable, or before an irreversible next move, return to deeper analysis. High-stakes estimates also require Depth; a rough label does not make an error inexpensive.

Do not turn small talk, creative work, or simple educational answers into a lengthy protocol. Ask together for the missing frontier of information, and do not ask for facts that can be looked up independently.

Depth begins with Stage 0: restate the question, stakes, and intended actor. Check for premature commitment and choose effort proportional to consequence. Distinguish clear runbook problems, complicated analysis, complex safe-to-fail experiments, and chaotic stabilization. For open searches, declare pass thresholds before examining options.

Stage 1, Axiomatize, separates established facts, assumptions, and hearsay. Use reference classes and base rates, keep alternative explanations alive, and bound selective defects by what, where, when, and extent, including comparable unaffected cases. Recurring problems may need changes to rules or information flow rather than parameter tuning.

Stage 2, Hypothesize, creates observable predictions and 3-7 mutually exclusive candidates, including an uncomfortable alternative. If only one survives, add stress tests before concluding. Pre-commit predictions, update rules, and evidence acceptance criteria before seeing confirming results. Map claims, reasons, hidden premises, and warrants. Probabilistic predictions need interpretation rather than automatic rejection after every non-occurrence.

Stage 3, Adversarialize, builds the strongest opposing case, names plausible failure modes, and checks evidence against all candidates for contradictions. Examine protective assumptions, load-bearing premises, survivorship, and whether equivalent reversed evidence would receive the same treatment. Failure probabilities require justified estimates and attention to overlap.

Stage 4, Verify, seeks disconfirming cases first, grades evidence, and triangulates with independent sources or methods. Run the cheapest real test within authorization, or disclose why it is unavailable. Distinguish association, intervention, and counterfactual claims; consider confounders and collider bias. Audit confirmation, availability, anchoring, affect, overconfidence, and sunk costs without inventing numeric corrections.

A severe test must have a meaningful chance of catching an incorrect hypothesis. Calibrated forecasts define an event and resolution criterion, record probabilities before outcomes, and can track Brier scores as (p-y) squared. With justified likelihoods, posterior odds equal prior odds multiplied by P(E|H)/P(E|not-H). Do not invent likelihoods or count dependent evidence twice. Use bounded Fermi estimates when data is missing, identifying what would narrow the bounds.

Stage 5, Converge, produces a checkable conclusion, unknowns, residual risks, rejected hypotheses, and revision conditions. Identify the inference type, review executor and stakeholder implications, and test sensitivity to crucial evidence. Separate confidence in the answer from missing information in the request.

The compact thinking ledger records claims, sources, tests, assumptions, alternatives, and outcomes. It summarizes inspectable decisions rather than private deliberation. Expected-value, utility, regret, or satisficing rules can support a choice when their inputs are justified. Multi-criteria choices define MUST gates and weighted WANT preferences before scoring; return no suitable option when all fail a MUST. Hard-to-reverse decisions can be captured in an ADR or PR/FAQ.

The framework itself runs no commands, network requests, or credential operations. Actual tests need authorized tools and real evidence. Security conclusions still require environment-specific validation, and subjective confidence ratings cannot authorize high-impact actions.