2. ACE and the neurosymbolic bargain

Section 1

Three ways Meridian gets hurt

Next section

Section 2

Deterministic and non-deterministic reasoning

Next section

Section 3

Dividing the work

Next section

Section 4

From data to wisdom, and the tier AWS added

Next section

Section 5

The seven standards you are about to meet

Next section

Section 6

What the lab proves

Hands on

Watch consistency break

20 min

Run one natural-language loan question through a non-deterministic scorer five times and get five answers, then run the same question through a deterministic SPARQL rule five times and get one.

make lab-up            # once, starts Postgres + Fuseki + Ontop
make lab lab02-ace-coinflip

Full instructions, starter files and the solution are in labs/modules/lab02-ace-coinflip/README.md. The lab checks its own work, so you will know if it passed.

Check yourself

Quiz

6 questions, pass mark 70%. Answers are explained either way, so a wrong guess is not wasted.

1 An agent computes Daniel Okafor's debt-to-income ratio as 0.34 when the correct figure under Meridian's policy is 0.41. The arithmetic is flawless. What went wrong?

The model needs a larger context window It divided by gross monthly income when the policy specifies net, and that definition was never written down where the agent could read it The income figure in the warehouse is corrupt Debt-to-income cannot be computed for self-employed borrowers

2 Setting a language model's temperature to zero gives you:

Full determinism, so the output is safe to put on a regulatory return Less variation, but no guarantee of reproducible output across time, versions or serving conditions Faster inference and nothing else The same behaviour as a SPARQL query

3 Which two tasks belong on the symbolic side of the neurosymbolic split? (pick two)

Select all that apply.

Extracting the employer name from a scanned payslip that spells it three different ways Proving that loan L-2087 cannot be classified as both a secured and an unsecured loan Confirming that application APP-9002 was declined under rule R5 rather than R4 Writing the decline letter to Tobias Fenn in plain English

4 The AWS guidance adds an Intelligence tier to the classic Data, Information, Knowledge, Wisdom hierarchy. What does that tier hold?

Faster query caching for the knowledge layer Meta-knowledge: continuous self-assessment of the ontology and graph, and proposals for semantic evolution The language model itself The business rules that derive new facts

5 Loan L-2087 is 47 days past due. Under Risk's rule, days past due at or above 90, a deterministic SPARQL ASK query returns:

true, because the loan was restructured false, and it returns false every single time it is run It depends on the model's temperature setting true, because 47 days is materially overdue

6 OWL and SHACL both appear to constrain data. What is the key difference?

OWL is a query language and SHACL is a data model OWL reasons open-world (missing information means unknown, not false), while SHACL validates closed-world (missing information is a reportable violation) SHACL runs on AWS and OWL does not OWL only works on relational data