LaunchStack

Transparency

Methodology

Where LaunchStack data comes from, how scores are computed, and how often they refresh.

Sources

LaunchStack is a public catalog of AI companies, model versions, and stack categories. Entity records are researched by our agent crew (Robin research → Quinn QA → publish) and enriched from provider documentation, public filings, funding announcements, product pages, and leaderboards.

Every claim that gates publish readiness is expected to carry a source URL and confidence level. Scraped or unverified rows are excluded from the default public directory unless you opt into “Include unverified.”

Momentum (companies)

Momentum is never invented. If a company lacks enough signals, the UI shows Not scored and the company is excluded from rankings, radars, and scatters.

Computed by refresh_company_momentum() into company_momentum_snapshots. Each available input is scored 0–100; the composite is the simple average of available inputs. A score is written only when ≥ 2 inputs are present:

  • Funding recency (latest announced round)
  • Headcount delta (≥2 employee_count observations)
  • Web-traffic delta (monthly visits series)
  • Model-release cadence (developer’s model_versions)
  • Event count (last 12 months)

Snapshots include an as_of timestamp and an expandable inputs breakdown on the company profile.

Value score (models)

Model value comes from curated score_observations for score_definitions.key = model_value. There is no default of 70 — unscored models render as Not scored and are excluded from value leaderboards.

Capability fields (context window, max output, pricing, knowledge cutoff, deprecation, API docs, and benchmarks JSON for MMLU, GPQA, SWE-bench, LMArena Elo) are backfilled from provider docs and public leaderboards with per-score source URLs and dates when available.

Refresh cadence

  • Catalog research — continuous via Inngest agent runs (Robin / Quinn / Scott).
  • Momentum — recomputed on demand via POST /api/admin/refresh-momentum or select refresh_company_momentum(); typically after metric or funding updates.
  • Model capabilities — curated backfill script / admin real-scores pipeline when provider docs change.

Coverage chips on list pages (e.g. “Sorted by momentum · 48 of 1261 scored”) show how much of the current filtered set has a real score for the active sort column.