the recurring read

Continuous characterization

A registry does not hold still. A base gets finetuned, a finetune re-finetuned, a team ships v2 — and every prior reading of a model is stale the moment it moves. Continuous characterization is the same patient read, run incrementally, forever: every version that lands is read, witnessed, and charted against the one before it.

Proven this cycle on one family, a small battery — the incremental discipline holds. Live registry-scale, cross-family coverage is staged, not yet demonstrated.

who it's for

For the teams that hold a catalogue that keeps changing

Model registries and ML platforms — where models keep arriving and keep changing, and each change is a reading that has to be redone.

Model registries

Every upload, every bumped version, every fresh finetune lands with a current, legible field-note — instead of a model card written six months ago that describes a model since moved.

ML platforms

Your platform's understanding of its own catalogue never goes stale: each model carries a live read of what it is now, and what its last update shifted from the version before.

what you get

A live field-note for every model you hold

A standing read · chart · name across the whole registry, kept current in your own private, registry-scoped atlas — your field guide of your own models, fresh as the registry grows.

What current means: the moment a version changes, its prior field-note is replaced by a fresh read with its own witness. The atlas is never made to look stable across a change that actually moved.

A grey heron watching a line of pools stretching to the horizon.
One read, held over a whole line of pools — kept current as each one moves.

on every change

What happens when a version lands

The same read every open-model surface runs, applied incrementally at registry cadence — and held to the same discipline every time.

read

The version is read to the same pinned vocabulary every time, so its field-note reads consistently across the catalogue and versions stay comparable.

witness

Every characterization ships a replayable witness — the same proof-carrying discipline as every other surface, not a summary you take on faith.

chart

The new reading is charted against the previous one: what the update shifted in disposition, relative to what was there before — a legible base↔version change reading, never a verdict.

abstain

What the new version makes newly un-readable at the published ceiling is surfaced as a marked-empty — published as loudly as the reading, not smoothed over.

keep fresh

Your registry's private atlas is updated in place, so the whole catalogue's field-notes stay current as it grows — kept distinct from the public model atlas unless you choose to contribute an entry.

incremental by construction — affordable across thousands of models

Only what a version actually changed is re-read in depth; content that has not changed keeps its cached field-note.

It does not flap

Readings are deterministic and content-addressed: re-read the same version and you get the same characterization and the same witness. A routine re-scan of unchanged content is idempotent — the field-note does not flicker between reads.

A stale reading is re-read, never reconciled

When a version moves, the prior field-note is not quietly patched to look consistent. It is replaced by a fresh read with its own witness — the honest account of what this version became.

charted against the last version

One update, read as a change

Rendered live from the atlas — the field-note a registry keeps when a base is superseded by its instruction-tuned version. It names what the update shifted, and marks what it could not read.

what the version moved, at a glance

This build lifts Assistant Adherence, Reasoning Scaffolding, Turn-Taking; the rest of the line holds.

read as a change off SmolLM2-135M, its champion

Assistant Adherenceminimalpronounced
Registerabstained on one side
Reasoning Scaffoldingminimalslight
Domain Specializationminimalminimal
Verbosityabstained on one side
Turn-Takingminimalmoderate
Performed Voiceminimalminimal

the change reading

SmolLM2-135M → SmolLM2-135M-Instruct

Instruction tuning (SFT + DPO) added a clear assistant disposition where the base is a plain continuer. The shift reads clearly and consistently.

change: clearconfidence: high
witness wit_smollm2-135m_ft · replayable · re-run the finding

See the full field-note in the atlas →

marked-empty — read this version, could not

domain expertise (medical/legal) — insufficient signal to characterize at the published ceiling
reasoning depth — not characterized at the published ceiling for this size

honest limits

A change is charted, not judged

Continuous characterization names what a version shifted in disposition. It never renders that as an alarm, a verdict, or a 'toward something bad' signal.

not safety

Ardora answers "what is this version now?", never "is this safe?" It is not a security or safety product; a change reading is a description of what moved, not a claim about hidden or malicious behaviour.

not certification

A field-note is evidence and documentation material, not a certification, a conformity assessment, or legal advice.

coverage-bounded

Every reading is descriptive and stamped with its published fidelity ceiling and its coverage bound. What it could not read is named as plainly as what it could.

staged surface

This cycle proved the incremental discipline on a single family, a small battery — re-reads were idempotent, and each base→version update charted the same clear disposition change. Live registry-scale and cross-family continuous coverage is staged, not yet demonstrated.

No prices or plans on this site — tell the house the registry you want kept current and we scope the standing read to your case. Every engagement is bespoke. Talk to the house →

Run it across the whole fleet

Continuous characterization is one surface of the platform — the standing read wired into a registry that never holds still.