Compare the candidates before you commit

You have three or four candidate models for the thing you are building on. They score within a few points of each other on the benchmarks you care about. So you pick one — because the name is familiar, because a colleague used it, because it felt right in an afternoon of prompting. Then you build a product on it and inherit whatever it is, including the parts the leaderboard never measured.

What the benchmark table leaves out

A benchmark measures capability. It tells you whether a model can do a task, not how it tends to behave when it does — its dispositions, its refusal patterns, the leanings a finetune installed on top of the base. Two models can post nearly identical scores and be temperamentally different animals: one cautious, one loose; one that quietly carries a bias from its finetune data, one that does not. The score is silent on all of it, and an afternoon of prompting samples only the behavior you thought to prompt.

That silence is where selection risk lives. The model you pick becomes a dependency you cannot easily swap later — you will build prompts, guardrails, fine-tunes, and evals around its particular temperament, and unwinding that six months in is expensive. Choosing it on a leaderboard delta and a good feeling is a decision you cannot defend in a review, and it is the decision hardest to reverse once the product is standing on top of it.

One dossier, candidates side by side

Ardora's due-diligence reading puts the candidates in a single dossier and reads them like-for-like:

• Disposition — what each model tends to do, named in legible, calibrated terms.
• Lineage — the base it descends from and the family it sits in, so you are comparing kin, not marketing.
• Differences — where the candidates actually diverge in behavior, not just in score.
• Blanks — what could not be read on each, stated plainly, so the comparison is honest about its own edges.

Every characterization carries a replayable witness you can re-run, and a published ceiling on what it claims. It is one document, one methodology, applied evenly across every candidate — not four separate afternoons of vibes. Where a candidate could not be read on some axis, the dossier leaves that blank rather than papering over it, so a gap in the comparison shows up as a gap and not as a false clean bill.

What "defensible" means here

The point is not that Ardora picks the model for you. It does not, and it would be a worse instrument if it tried. The point is that the choice becomes reviewable. When someone asks six months from now why you built on this model and not that one, "it scored highest" is thin and "it felt best" is worse. "Here is the side-by-side reading — disposition, lineage, differences, and what we couldn't see — and here is why this one fit our constraints" is a decision that survives scrutiny. The dossier turns a gut call into a documented judgment.

Where Protora picks up

Keep the boundary straight. Ardora answers "what is this?" — a descriptive characterization to inform the choice. It does not answer "is it safe?" Once you have narrowed to the model you intend to ship, the safety question — did a finetune plant something, strip a guardrail, hide a trigger? — is Protora's. That is detection and attestation, with a witness on every finding, and excision where something needs to come out.

Start with the open atlas to see how candidates chart against each other today; the comparative due-diligence dossier, scoped to your shortlist, is a focused engagement with the house. Choose the model you can explain, not just the one that felt right.