AI Tools Index / Concepts
concepts · 8 explainers

The vocabulary you need before you buy anything.

This section explains the machinery behind the index. The model reviews say what we think of a particular set of weights; these explainers say what the words mean — calibration, abstention, effective context, quantisation — and why they decide whether a deployment is safe for client work. Each one is written for the person who has to sign the policy and answer for it: a COLP, a managing partner, an IT lead fielding hard questions from the partnership. Our vocabulary, plainly, with the uncertainty left in.

8 min read

What calibration means for legal work

Why do two models that score the same on a benchmark behave completely differently on a lease?

read the explainer →
8 min read

Abstention: why refusing to answer is a feature

How do we stop a model inventing an answer when the file does not contain one?

read the explainer →
7 min read

Temperature, top-p and why zero is not always the answer

Which settings should we standardise across the firm?

read the explainer →
7 min read

Advertised context versus effective context

It says a million tokens, so why did it miss the clause in the middle?

read the explainer →
9 min read

Retrieval, fine-tuning or a better prompt?

The model keeps getting our work wrong — what do we actually fix?

read the explainer →
8 min read

Reading an open-weight licence

What does 'open' actually let us do with client data?

read the explainer →
8 min read

Quantisation: what you lose when you shrink a model to fit

Can we run a serious model on hardware we already own?

read the explainer →
9 min read

Building an evaluation harness from your own matters

How do we know whether any of this works before we roll it out?

read the explainer →
how to read these

Each one answers a question a COLP actually asks.

They are written for the person who has to sign something, not the person building it. Where a concept is genuinely uncertain, we say so rather than inventing a rule.

then test it

Reading is not evaluation.

The only way to know whether a model abstains on your files, or quotes your precedents faithfully, is to run it on your own matters with your own fee-earners scoring the output.

Start with a readiness assessment →