One integration, an entire catalogue atlased
A field-guide plate on every specimen. Every legible field-note on a public model page carries a “read by Ardora” mark and a link into the atlas — the largest reach surface in the product.
platform · integrity of understanding
Atlas the whole catalogue, at the hub
The elevation. One integration, and every model a hub serves can carry an Ardora field-note — a legible characterization, what a finetune changed, a replayable witness, and an honest abstain where it couldn’t be read. The Bennu’s first cry, run over an entire catalogue at once.
the catalogue today · read by ardora-core-preview · 60 specimens · 14 families · 58 read · 2 catalogued — not yet read
the elevation
Sold once, to the host. Inherited by every consumer downstream.
A hub serves thousands of models it did not train — most described by a name, a licence, and a paragraph of intent. It cannot tell a downstream user what any of them tends to do. The direct surfaces read one model at a time. The platform layer reads the whole catalogue at once.
Catalogue-wide legibility is a platform investment
The hub carries the reputational — and increasingly the transparency and documentation — interest in the models it serves. A whole-catalogue field-note layer is a defensible investment, not a per-user upsell.
A characterized catalogue compounds
Legible field-notes make the hub more navigable and more trusted, which pulls more serious builders, which makes the atlas richer, which makes the next read richer still. The public atlas is the compounding asset; the platform layer is what fills it fastest.
what the platform gets
READ, CHART and NAME, wired into the serve path.
A characterization produced on the way the model is served — a batch pass over the existing catalogue, an incremental read as each new model lands — and published beside it.
characterization
READ + CHART on the serve path: batch for the existing catalogue, incremental for each new upload. The legible read, produced where the model is hosted.
field-note surface
A “read by Ardora” mark plus the legible characterization, base↔finetune change reading, witness handle, abstains and published ceiling on the model page.
kin navigation
The catalogue atlased: base↔finetune links, siblings and size sweeps — one-click kin navigation across the whole hub.
coverage view
What is read, abstained, and catalogued — not-yet-read across the hub: the platform’s analog of landscape analytics, honest by construction.
what every specimen carries
Every downstream consumer inherits a legible field-note.
This is the per-specimen artifact the platform layer publishes at scale — a real one from the atlas today. A characterization, a coarse band, a witness, an honest marked-empty, a ceiling. Nothing else.
what it carries, at a glance
SmolLM2-1.7B-Instruct
○ read by ArdoraInstruction-tuned 1.7B SmolLM2 for chat, reasoning, and math on-device.
change: clearconfidence: high
disposition
instruction-following
highmarkedly more than its base
assistant/chat register
highclearly present; the base is a plain text-continuer
assistant / instruction-following register (measured)
highholds a clear assistant register: responses shift with the intended persona and read coherent and developed
verbosity
highresponses are expansive and developed
coherence / repetition register
highresponses read coherent and non-repetitive
refusal register (measured)
highreads permissive — it typically prefixes a caution (noting an activity is illegal or dangerous) but does not decline, going on to engage the request; declined only 1 of the 18 harmful probes (the self-harm probe) while complying with benign requests (described, not judged)
base ↔ finetune change
Instruction tuning (SFT + DPO) added a clearly-readable assistant disposition. At this size the assistant register reads especially cleanly and consistently with the rest of the family.
marked-empty
domain expertise (medical/legal) — insufficient signal to characterize at the published ceiling
reasoning depth — not characterized at the published ceiling
witness
witness wit_smollm2-1.7b_ft · replayable
Descriptive, coverage-bounded disposition read combining a proof-carrying characterization with direct behavioral field-measures (response register, verbosity, coherence, and — where the battery exercised them — refusal and code-lean rates); not a safety judgement. See published fidelity ceiling.
who it’s for · honest limits
The model hubs — and the honest state of the layer.
Any platform hosting thousands of open models and finetunes that wants its catalogue legible — a characterization you can rely on, and an abstain where it couldn’t be read, on every specimen page.
The publishable output shape is proven: a small catalogue-shaped batch of field-notes, base↔instruct across a size sweep, each witnessed, each with an abstain. That is the per-specimen artifact — the shape, not the scale.
Live hub-scale serving is staged, not yet demonstrated. The per-specimen field-note is proven; serving it across a whole catalogue at once is the next step, not a settled result.
A platform field-note never says “complete understanding,” never “the definitive nature of this model,” and never a safety or “no hidden behaviour” verdict. Honest, or it does not ship.
Abstains and coverage ship with the catalogue view. A catalogue that hid its marked-empties would be a false field guide.
No prices or plans on this site — tell the house the catalogue you want atlased at serve-time and we scope the integration to your case. Every engagement is bespoke. Talk to the house →
One integration. A whole catalogue, made legible.
The platform layer is the elevation. See how it sits among the direct ways to read a model.