The interface selector.
Five requirements in, five interfaces out. The ranking is a documented rule table, not a learned model: change one input and the reason changes. The point is to make the design space explicit, not to pick for you.
The interface selector
Ranked shortlist
Points are rule weights, not probabilities, benchmark scores, or a recommendation. The bar is relative to this shortlist.
How the score was built
Every row is a rule that fired. The total is the sum of the rows, and the ranking is the total with a fixed tie-break.
Five inputs, a small table, no hidden model.
The selector adds points from five rows: one base row for the data type and one row for each of the other four inputs, plus a fixed tie-break. The tables below are generated from the same object the selector uses.
They are not learned weights, not calibrated probabilities, and not estimates of success. They encode the lecture’s qualitative trade-offs so that the ranking can be inspected. A different team with different costs should change the weights — and then defend the change.
The shortlist moves with the requirement.
The design space is small
Five interfaces, five inputs, and a handful of interactions. Most of the argument about a neurosymbolic design is an argument about which row should dominate — not about which brand to use.
Requirements trade off
A required guarantee pushes the verifier and search up and the soft loss down. A low budget pushes search down. An untrusted symbolic model pushes hard constraints down. The movement is the lesson.
It is not a benchmark
The points are not measured. The selector knows nothing about your data, your latency, your team, or your failure costs. It cannot tell you which shortlisted design will work.
It is not a recommendation
A high score is a hypothesis about fit, not a verdict. The honest output is a shortlist plus a reason, and the next step is an ablation, not a commitment.
Three settings worth trying
- Perception · weak · required · low · no. A hard guarantee is wanted, the symbolic model is suspect, and there is no budget for search. Read the trace: which interface survives, and what does the reason admit?
- Language · none · required · high · yes. The classic proposal-and-check setting. Search and verifier should rise; compare their reasons and their trade-offs.
- Graph · full · not needed · low · yes. The structure is the supervision and the deployment is cheap. The data interface should dominate; ask what it still cannot guarantee.
Back to the selection procedure · Read the guarantee taxonomy again · Open the slides