A disciplined protocol for attempted refutation.
The method is designed to make it hard for the machine to fool itself, and impossible for it to grant itself authority it has not earned.
Principles
Seven commitments.
- 01
Frame before modeling
Event, horizon, population, admissible assumptions and materiality are defined before any model is selected.
- 02
Interrogate with heterogeneous engines
Statistical, stochastic, symbolic, fuzzy and learned families examine different properties; none certifies itself.
- 03
Fix the perturbation set in advance
Windows, frequencies, resamples, seeds, specifications and regimes are declared before results are seen, to avoid a convenient multiverse.
- 04
Qualify, then synthesize
Only candidates classified as persistent or conditional enter synthesis, and classifications remain revisable.
- 05
Discount dependence, preserve conflict
Shared data and inductive bias reduce evidential weight; informative disagreement is retained in the artifact.
- 06
Calibrate and expire
Probabilities are calibrated against outcomes and carry a validity horizon after which they are stale.
- 07
Keep the modeler accountable
Automation may propose structural revisions. Promotion requires defensibility checks and explicit sign-off.
Conditional complexity
Complexity must pay for itself.
L0–L9
The architecture the principles govern.
Computational epistemology
Not the most elaborate machine. The smallest one capable of defending its inference.
From the working paper · governance principle
Beyond Ontology: Computational Epistemology as a Layered Inference Machine
Fernando Negrini et al. · Working Paper v1.1 · 2026
Ontology structures what is available to the machine. Computational epistemology structures what the machine is permitted to conclude from it.