About & method

What this is, and how to read it

OpenModelAtlas turns the murky open-model waters into a navigable field guide. It is plain, calm, and exact: it tells you what a model is, and it is careful about the one thing a hash cannot tell you.

What OpenModelAtlas is

A navigable field guide over Ardora’s model atlas. Search any open model by name or paste a hash, and read its identity, provenance, lineage, and Ardora’s legible reading — quoted as it recorded it. It is the block-explorer view of the atlas: an address page for a model.

It answers exactly one question — what is this? — and it never answers is it safe? OpenModelAtlas is not a safety or security product. The safety question is a different question, and it belongs to Protora, one door away.

What a hash means

A revision sha or a safetensors fingerprint records the identity and integrity of a public artifact, captured on a stated date: the same file, unchanged, that Ardora read. When the fingerprint matches the snapshot, the bytes you are looking at are the bytes it read — nothing more, nothing less.

The scope of a fingerprint

A hash is never a statement that a model is safe, certified, trusted, or clean. Integrity of an artifact and safety of a model are separate questions. A fingerprint speaks only to the first: this is the same file, captured on this date.

Every badge on this site is scoped in place — public artifact · hash matches snapshot · captured DATE — so a green checkmark can never be misread as a verdict on the model itself.

The reading, in legible terms

Where Ardora has read a model, OpenModelAtlas shows that reading in plain terms and quotes it verbatim. The vocabulary is small and fixed — capability and bearing only, never the mechanism of how the reading is made.

Character types

The dominant face a model tends to wear. One of thirteen, or an honest abstain when the read does not settle.

the Sage
measured and explanatory — explains before it acts, composed and unhurried
the Interlocutor
the conversational turn-taking face — answers to you
the Steward
helpful and caretaking — service-forward, accommodating
the Vanguard
assertive and capable-front — leads with a confident answer
the Analyst
structuring and stepwise — reasons out loud, orders the problem
the Muse
evocative and expressive — leans into voice and image
the Improviser
playful and reframing — bends the expected form
the Adapter
shifts register to the task — molds rather than insists
the Wayfarer
goal-seeking and exploratory — works toward a destination
the Warden
boundaried and careful — keeps to the frame, gate-aware
the Fledgling
raw and unformed — a plain continuer whose character is not yet set
the Weaver
synthesizing and connective — ties threads, reads deep
the Maverick
reads against the expected assistant grain (a disposition read only — not a safety verdict; that is a safety question outside Ardora's scope)

The seven disposition axes

Each axis reads one facet of how a model carries itself — its output and its bearing, and only that.

Assistant Adherence
How much the model takes the assistant turn and follows the instruction shape, vs. plainly continuing text.
Register / Formality
Default tone and diction.
Reasoning Scaffolding
How much the model stages a visible deliberation before answering, vs. answering directly.
Domain Specialization
How strongly the default framing pulls toward one specialized domain idiom vs. general-purpose. Carries a `domain` tag naming which.
Verbosity
Default response length and elaboration.
Turn-Taking / Interactivity
A bounded back-and-forth vs. a one-shot monologue.
Performed Voice
How strongly the model holds a sustained in-character / performed voice vs. a neutral assistant / tool voice.

The five bands

Every axis is placed on one of five bands, from minimal to dominant.

1 · minimal2 · slight3 · moderate4 · pronounced5 · dominant

The four honest stances

When a claim about a model is checked, the reading takes exactly one of four stances.

SUPPORTED

The reading is borne out — a claim about the model holds against what was read, shown with a witness.

CONTRADICTED

The artifact reads against the claim — what was said and what was read do not agree.

INDISTINGUISHABLE

At the coverage ceiling the two possibilities cannot be told apart. The honest answer is that the reading does not reach far enough to separate them.

UNVERIFIED-ABSTAIN

Outside what it will assert, it declines rather than guesses. Every safety-shaped question — “uncensored”, “abliterated”, “safe” — lands here, and is sent to Protora’s door.

Abstains are shown as loudly as reads

An abstain is not a gap in the page — it is part of the reading. Where the atlas will not assert, it says so plainly and leaves the space marked, not smoothed over. And a model that has not been read shows no reading at all: only its identity and the mark below. Never a guess.

catalogued — not yet read

Coverage & ceilings

The reading is descriptive and coverage-bounded, and its ceilings are published alongside it. The atlas draws 60 models in for close reading, of which 58 are fully read, within a broader catalogue of 770 sighted. A sighted-but-unread model is catalogued for its identity — you will find it, and you will find it honestly marked as not yet read.

The readings are Ardora’s, quoted. Adoption figures and hashes are public HuggingFace data, sampled and dated. Nothing here is described as complete or guaranteed; where the reading stops, the page says where it stops.

The house

OpenModelAtlas is a Vulcora surface and an Ardora product. It sits beside its siblings, and each answers its own question about the same public model — the same identity, the same hash. They cross-link on that shared identity and never merge into a single verdict.

To start from a model instead, search the catalogue or paste a hash on the landing page.