read · self-serve

The free read: name a model, get its field-note

Pick an open model from the atlas — or pour your own in — and get a legible reading of what it is: how it tends to behave, and, where a base is named, what its finetune changed. Each finding carries a replayable witness; where the shape can’t be seen clearly, it abstains.

full witnesses · full abstain honesty · full fidelity to the published ceiling · capped on volume, never on trust

who it's for

The builder deciding what to run

The individual ML engineer, the curious builder — anyone standing at the edge of the open-model waters asking the plain first question before they commit.

It turns “what does this model actually do?” into a habit — one legible field-note at a time.

One clear pool, one stone, one visible fish — a single model read on demand.
One pool, one stone, one visible shape: the atlas's per-model field-note, produced on demand for a model of your choosing.

a real field-note, at a glance

SmolLM2-360MBase360Mscoreminimal

Name an open-weight model, optionally name its base or a sibling to place it against, and Ardora returns the field-note: the legible reading, the per-finding witnesses, the family placement, and the abstain list. Nothing is named that could not be clearly seen.

It is the same read that fills the public atlas — the wedge, not the ceiling. Every entry a builder publishes back widens the field guide and draws the next builder in.

what you get

The field-note, in full

Six things, every read — or an honest blank where one couldn't be seen.

the disposition read

A legible characterization of how the model tends to behave — witnessed, and carrying its confidence band.

the finetune-change read

Where a base is named, what the finetune shifted from it — in legible terms, with a coarse magnitude band and a confidence.

family placement

Where the model sits among its kin, and the plain-language difference from its base or a sibling.

per-finding witnesses

Click any part of the reading to see its replayable artifact — the proof that the characterization reproducibly holds.

the abstain list

Every trait Ardora would not characterize, each marked with a reason — shown as loudly as the reading itself.

the published ceiling

Every read carries its confidence band and the engine version, stamped with the published fidelity ceiling — never exceeded.

the discipline, made concrete

Read. Calibrate. Or abstain.

Three plates that hold across every read — grounded here in a real entry from the atlas seed.

Read

A finding is demonstrated by a replayable witness — the characterization reproducibly holds, not a claim you have to take on faith.

witness wit_smollm2-360m_base · replayable

Calibrate

Every reading carries a confidence band. Nothing is presented as certain that isn’t.

text continuationconfidence: high

Abstain

Where the shape could not be seen clearly, Ardora says so, with a reason — never a manufactured reading to look thorough.

domain expertise (medical/legal) — insufficient signal to characterize at the published ceiling

Drawn from SmolLM2-360M in the atlas — read, not asserted.

free — and where it's capped

Capped on volume, never on trust

The self-serve read is the way in — the adoption and content engine, not the thing we sell. So it is capped where cost lives, not where honesty does.

The self-serve read

Full witnesses. Full abstain honesty. Full fidelity to the published ceiling — capped on volume and model size, never on trust.

Free to run — capped where cost lives, never on honesty.

Deeper reach

Larger models, more runs, saved bases and comparisons, priority processing — for the builder who lives in the atlas.

No prices or plans here — scope the deeper reach with the house. Talk to the house →

The deeper questions this read raises live up the ladder: which candidate to build on, what a finetune actually changed, and keeping every reading current.

honest limits

What the free read does not do

Owned out loud, because the discipline is the point.

not a leaderboard

It is not a benchmark or a ranking. It reads disposition, not a score.

not a platform

It is not a fine-tuning platform, a model host, or an inference API.

not a chatbot

It is not a chatbot-about-your-model. It hands you a field-note, not a conversation.

not a safety check

It does not answer "is this safe?", flag hidden behaviour, or clear a model. It reads what a model is — never whether it is safe.

where it stands — proven this cycle, deliberately modest

On a single-family seed — SmolLM2, a Llama-family size sweep — the read names the same legible disposition consistently: a clear assistant disposition added by instruction tuning, in the same direction across the size sweep, each finding carrying a replayable witness, each read publishing its abstains rather than reaching past what it could see. The self-serve portal is on the roadmap; the read it will serve is running in the atlas today.

60 specimens · 14 families · 58 read · 2 catalogued — not yet read

No prices or plans on this site — tell the house what you need read, a single model or a deeper reach across many, and we scope it to your case. Every engagement is bespoke. Talk to the house →

Every read starts as a field-note in the atlas

Browse the open-model waters, one legible entry at a time — and see the shape of the read you'd get.