The base↔finetune diff, with evidence
You trained for a change and got back a new set of weights. What actually moved between the base and the finetune is the thing no model card can tell you — and the thing this reading is for.
The gap between intent and result
A finetune is a hypothesis. You picked a base, assembled a dataset, chose a recipe, and ran it because you wanted the model to acquire some behavior — a tone, a format, a domain skill, a refusal policy. What you get is a new set of weights. Somewhere between the intent and the weights, the model became something, and the only record you usually have of that becoming is the sentence you wrote in the model card.
That sentence describes what you trained for. It does not describe what the model became. Training moves more than the thing you were aiming at — it nudges dispositions you were not watching, in directions you did not choose, by amounts you cannot eyeball. The card goes stale the instant the run finishes, and nothing in the file marks the difference.
What the diff actually shows
Ardora reads that becoming directly. Given a base and a finetune off it, it produces a base↔finetune diff — a like-for-like reading of what the training actually moved:
- Which behaviors shifted — not the ones you asked about, the ones that changed.
- In which direction — toward or away from a disposition, named in legible terms.
- By how much — a stated magnitude, with confidence, not a vibe.
- What it could not read — the honest blanks, published as plainly as the movements.
Every entry ships with a replayable witness: a reproduction you can re-run to see the shift for yourself. A reading with no abstains deserves a raised eyebrow — Ardora publishes the marked-empty as loudly as the movement.
Material that doesn't go stale
This is the raw material for the documentation that always rots first. A model card that says "finetuned for customer-support tone" is a claim. A base↔finetune diff that says these three dispositions moved this far in this direction, here is the witness, here is what stayed dark is evidence — the kind you can drop straight into release notes, a model card, or an internal review without hand-waving.
It is also a second pair of eyes on your own run. When the diff shows a disposition moving that you did not intend to touch, that is not a failure of the reading — that is the reading doing its job. You get to decide whether the drift is acceptable before it ships, instead of discovering it in production.
Descriptive, not a verdict
Read this boundary carefully, because it is the whole design. Ardora is a naturalist's instrument. It answers "what is this?" — it does not answer "is it safe?" A base↔finetune diff is a characterization of disposition: here is what moved, here is the evidence, here is the ceiling on what could be read. It is coverage-bounded and never absolute.
The safety question — is this movement a planted backdoor, a stripped guardrail, something you must cut out? — belongs to the other house. That is Protora's detection and attestation, and where something is in the model that should not be, excision with proof that only that changed. Two houses, two questions; Ardora tells you what the model became, and hands off cleanly when the question turns to whether you can trust it.
Where to start
The open atlas already charts finetunes against their bases where both are read, so you can see the shape of a diff before you commission one. The deeper base↔finetune-change reading — on your model, with your base — is a focused engagement with the house. Bring the two sets of weights; get back a diff you can put your name on.
