the product line
Ardora is one product — three verbs on one engine. It reads a model's disposition and what a finetune changed, charts it among its kin, and names the reading in legible, calibrated terms. The same instrument, surfaced as a ladder that climbs from open credibility to depth scoped with the house.
READ · CHART · NAME — three verbs, one engine, a ladder of depth
the ladder
The ladder is the spine of the product. It climbs left to right — earn trust in the open, open the first depth read, keep the readings standing as the catalogue grows, then scale across a portfolio and onto the platforms that host the models. Every rung is the same instrument, read at a different reach.
READ · CHART · NAME — a champion and one build off it, read in the atlas today
The base is the champion; its instruction-tuned sibling is a build off it — the little bars are the stat line, the glyph is the disposition class. This is the reading every rung below is a different reach into.
Every open model, read and set among its kin — free to browse.
for everyone — the market
An open, family-navigable field guide of characterized open models — read in the open, abstains and all. The credibility engine the whole ladder leans on, free by design.
Name a model, get its stat line and build story back — free.
for ML engineers · the curious
Pour in or pick an open model and get a legible reading — its disposition, what a finetune changed from base, a replayable witness, and an honest blank where it couldn't see clearly.
Your shortlist, read side by side — so the choice rests on evidence.
for teams choosing a model to build on
A comparative selection dossier across candidate models — what each tends to do, set side by side, witnessed, with the abstains marked. A decision, not a leaderboard.
What your build actually moved off its base — witnessed.
for model developers · finetuners
"What did my finetune change from base?" — a legible base↔finetune diff, witnessed, with its confidence stated and its ceilings published. The wedge reading, set into the atlas.
Every model's field-note kept current as the catalogue grows.
for model registries · ML platforms
Keep every model's reading current as the catalogue grows — readings that don't go stale the moment someone finetunes. Family and lineage charting across the whole set.
A witnessed, file-ready account of what a model is and what changed.
for AI-governance · model-risk functions
Evidenced, legible model-documentation material — a witnessed reading of what a model tends to do and what a finetune changed. Documentation material, not a certificate.
How a whole family or portfolio differs in disposition, charted.
for analysts · procurement · researchers
Disposition analytics across a family, a portfolio, or the whole ecosystem — which models carry a current, witnessed reading, and how siblings differ from one another.
One integration — every model you serve carries a field-note.
for model hubs · platforms
One integration atlases the whole catalogue at serve-time, so every downstream consumer inherits a legible, provenance-carrying field-note — without running Ardora themselves.
Earn the trust in the open first — then convert it into depth.
the frame
Ardora is complementary, not a replacement. Keep your eval harness for benchmark coverage and your registry for lineage bookkeeping. Add Ardora for what neither gives you — a legible, witnessed characterization of what a model tends to do.
A human-readable characterization of a model's disposition, and of what a finetune changed from its base.
Every characterization ships a replayable witness — a record you can re-run, not an opinion you take on faith.
It abstains rather than manufacture a reading, states its confidence, and publishes its ceilings — often modest.
The open model atlas: family-charted and lineage-aware, and it compounds — each read makes the next comparison richer.
Reads open-weight models — and, under contract, your own — without provider cooperation.
It answers "what is this?" — a naturalist's read of what a model openly tends to do.
A benchmark score is capability, not disposition — and it says nothing about what a finetune changed.
Not a confident explanation you can't verify. If it can't be demonstrated with a witness, it isn't recorded.
No "complete understanding," no "definitive nature of the model." It names only what it saw.
Not a report with nothing behind it. Every reading is set into a field guide that compounds.
No cooperation-required, provider-gated service. Open weights are enough.
It never answers "is this safe?" — not safety, not security, not certification. A different bird's water.
A plain probe can tell a finetune changed a model too. What Ardora sells is the disciplined, legible, navigable reading.
the order
Credibility before the ask; the cheapest proven thing first; depth where the commitment is. The order is not a funnel trick — it is the discipline, made into a sequence.
Credibility before anything is asked. The public atlas proves Ardora reads honestly — in the open, abstains and all — before anyone is asked for a thing.
A self-serve read of an open model needs no customer data and no integration. It is the free wedge that turns atlas readers into users.
Due-diligence, finetune-change readings, and continuous coverage are the committed work — decisions and obligations, not curiosity.
Every surface climbs on its own earned trust and its own infrastructure. The ladder is not shipped all at once.
Step onto the first stone. The atlas is free to browse, every reading carries a replayable witness, and every entry is a shape the heron actually saw — with the abstains marked in plain sight.
