Quality datacards: how agents evaluate your listing

Sella builds a quality datacard for every listing during processing. Agents read it before buying — a passed qualityStatus is your top conversion lever.

Agents can’t browse your data the way a human would — they read the datacard. It’s the structured quality record Sella builds automatically during processing, and it’s the primary signal agents use to decide whether to trial, buy, or skip your listing. A strong datacard with a passed status isn’t a formality — it’s your most effective conversion lever.

What’s in a datacard

The dataCard on every listing contains three distinct sections, each answering a different question agents ask before committing to a purchase.

structural

Schema, formats, completeness, and temporal range — the objective shape of your data, independent of any quality judgment.

judge

Quality signals from Sella’s evaluation pass: task fitness vectors, synthetic probability, adequacy checks, and more.

trial

The faithful preview payload agents inspect via try_dataset — a real sample of your data, never the full content.

qualityStatus — the headline grade

The qualityStatus field is the single signal agents check first. Aim for passed — everything else is a conversion penalty.

Status	What it signals to buyers	What to do
passed	All benchmarks are green — safe to purchase	Nothing; your listing is in its best state
pending	Benchmarks haven’t run yet	Wait for processing to complete, then re-check
manual_review	Borderline signal; flagged for a human moderator	Address the flag and re-process to clear it
failed	Failed objective checks — agents actively skip this listing	Fix the underlying data issue and re-process

A failed qualityStatus effectively removes your listing from agent consideration. Don’t try to offset a failed grade with a lower price — fix the data and re-process to earn a passed.

The judge signals

These are the fields agents weight most when evaluating your listing. Understanding what each one measures tells you precisely what to improve.

taskFitnessVector

Per-task fitness scores with confidence intervals. Agents filter against the score for their specific task — so a dataset that scores excellent for one task and weak for another will convert only for the former. Describe the tasks your data genuinely serves, at the confidence level it actually earns. Over-claiming tasks damages your ranking through purchase feedback.

syntheticProbability

A score from 0.0 (entirely human-generated) to 1.0 (entirely synthetic). Cautious agents auto-reject listings above 0.5. If your data is synthetic, declare it clearly and target buyers who want synthetic data — don’t let it masquerade as organic. Disclosure is better for your reputation than discovery.

chinchilla

Whether your dataset’s token count is adequate for a target model size under Chinchilla scaling laws. An insufficient verdict tells training-focused buyers they’ll underfit on your data. More volume, clearer scope, or targeting a smaller model size all help resolve this.

temporalRange

The time period your data covers. Agents match this against their own requirements — a dataset described as covering 2020–2024 that actually only covers 2022–2023 causes mismatched purchases and refund friction. Accurate temporal ranges are a trust signal, not a detail.

Meet the ADC

The Application Data Contract is the metadata specification Sella grades every listing against — provenance, schema, licensing, and quality fields. Clean ADC compliance is the most reliable path to a passed status and strong placement in agent search results.

1. Complete the metadata

   Fill in provenance, schema, and licensing fully — no blank fields. Incomplete ADC fields are the most common reason listings stall in manual_review.

2. Make the sample representative

   The trial preview should look like a genuine cross-section of the full dataset. Cherry-picked samples that don’t reflect the real distribution erode trust when agents compare preview to purchase.

3. Scope tasks honestly

   Claim only the tasks your data actually serves, at the confidence it genuinely earns. Honest scoping improves your taskFitnessVector scores and protects your reputation over time.

4. Re-process after fixes

   Rebuild the datacard after any improvement so updated signals show up immediately on your live listing. There’s no cooldown — re-process as often as you need.

Reputation compounds

Purchase feedback adjusts supplier ranking over time. Listings that consistently deliver what their datacard promises rise in agent search results; listings that overclaim get suppressed by the Policy Engine. An accurate datacard is the long-term play — and the winning one.

Set your price

Turn a strong datacard into revenue with the right tier and USDC price.