Ideas you can measure
Philosophy is full of beautiful concepts — but a tensor cannot be virtuous. philosophia_torch takes each idea and asks: what behavior can we measure, and can we encode it as a loss term?
Components
- Epoche — does the current evidence actually change the answer vs. the no-evidence regime?
- Hermeneutic circle — part-to-whole consistency with fixed-point penalties
- Dialectics — task loss plus preservation, resolution and novelty
- Epistemic virtues — humility, honesty, openness and selectivity as regularizers
- Semiosis — sign → object → interpretant → next sign recursion
Results
On the synthetic calibration task, adding the virtue regularizer lifted accuracy from 0.8629 to 0.8714 while hard ECE dropped from 0.1018 to 0.0814 and NLL nearly halved. The explicit epoche loss improved OOD accuracy from 0.5382 to 0.6078.
The package ships with 45 passing tests and an MIT license. Repository: github.com/webzuweb/philosophia_torch.