Methodology

How buildagentic measures things

Every figure on this site carries three things: the source that reported it, the moment it was observed, and how it was obtained. Nothing is averaged across sources, nothing is filled in when a sync fails, and no language model writes a displayed value. 616 values are served across 3 datasets, resolved from 14,856 observations on file — nothing is ever overwritten, so every correction we have made is still here alongside the value that replaced it.

Provenance

Four ways a figure gets here

Each method claims something different, and the difference is the point. A tooltip on any value names which one applies to it. The counts are values CURRENTLY SERVED, not rows on file — a superseded observation is history, not a claim.

synced

453 served

Read from a source API by an adapter, on a schedule. The figure is whatever the source returned at the moment shown, unaltered.

seeded

163 served

Written by hand when the dataset was created, from published documentation. Nobody has checked it since. It claims accuracy at the time of writing and nothing about today.

curated

0 served

Read and confirmed by a named person, whose name is stored with the value. This is the only method that claims review — a review nobody can be asked about is not a review.

derived

0 served

Computed from other observations on this site. It carries the freshness of the oldest figure it was computed from, not its own.

Nothing on this site is curated yet. The mechanism exists and the review axis is enforced, but no value has been read and confirmed by a person. Every figure here is either synced from a source or seeded by hand and unreviewed — treat it accordingly, and follow the source link before quoting a number that matters.

Freshness

What an old figure does

Age is a per-field state, not a per-page banner. Two values in one row can disagree about how current they are, and the page says so rather than averaging them.

Fresh

Observed inside the dataset's SLA. Shown plainly.

Aging

Past the fresh window, inside the stale one. The value is served and carries its age, because a figure a day old is usually still the best answer available.

Stale

Past the stale window. Still served, still marked — removing it would replace a known-old number with no number, which is worse. The badge is the warning, not a reason to hide the value.

Quarantined

The run that would have produced a newer value failed validation and was rejected whole. The previous good figure keeps serving and says it is quarantined. This is the rule the pipeline is built around: a failed sync never replaces good data.

Unknown

No observation on file that this page can serve. It renders as an em dash, never as zero.

Refusal

What we refuse to publish

A pipeline that always writes something is a pipeline that will eventually write anything.

Validators run before anything is written

Every incoming row is checked against per-field rules — type, range, and plausibility against the previous observation. A run whose failures cross the threshold is quarantined in full rather than partly applied, because a half-written sync leaves a dataset agreeing with neither the source nor itself.

Withheld is not zero

A value that fails validation is withheld and named as withheld. It is never plotted as zero and never averaged away. A chart that looks complete and is not is worse than one that says what is missing.

Retirement is a state, not a deletion

An entity that leaves a source is marked retired. Its page answers 410 Gone rather than 404, because the thing existed and we are saying so. Its past observations stay on file and keep appearing in history — they were true when they were recorded, and deleting them would rewrite the record rather than update it.

No language model writes a displayed figure

Models are used for classification and enrichment behind the pipeline. No model output is ever written to an entity field or a displayed metric.

Review has a horizon

A curated value goes overdue after 180 days and says so. Freshness asks when the value was observed; review asks when a person last looked at it. They are separate axes, and a value can be current on one and overdue on the other.

Sources

Where each dataset comes from

Read from the registry, not written here — a source added to a dataset appears in this table with no edit to this page.

Agent engineering frameworks

18 tracked · fresh under 48h · stale past 336h

54 of 308 served values here were written by hand and never reviewed. These are the fields no source publishes — the ones a person decided, like what a project is for. Until that changes they stay marked seeded rather than dressed up as measurements.

  • GitHub REST APIEvery 24 hours
  • npm registry downloadsEvery 24 hours
  • PyPI registry metadataEvery 24 hours
  • pypistats.org downloadsEvery 24 hours

Language and embedding models

18 tracked · fresh under 24h · stale past 240h

54 of 193 served values here were written by hand and never reviewed. These are the fields no source publishes — the ones a person decided, like what a project is for. Until that changes they stay marked seeded rather than dressed up as measurements.

  • Hugging Face HubEvery 24 hours
  • OpenRouter model catalogueEvery 24 hours

Local inference hardware

10 tracked · fresh under 720h · stale past 4320h

55 of 100 served values here were written by hand and never reviewed. We looked for a source to replace them and did not find one: the publishers that hold these figures either forbid automated use in their terms, do not answer, serve their specifications only to a browser, or publish product families in a form where a number cannot be tied to a specific model. Until that changes they stay marked seeded rather than dressed up as measurements.

  • Wikipedia infoboxEvery 7 days