Transparency

How Scoring Works

Every score on Launch Sentinel is derived from a documented, reproducible algorithm. This page describes the model exactly as it runs — weights, thresholds, and all known limitations included.

1. Two-Index Model

Each product receives two independent scores that are blended into a composite.

Profile Maturity (0–100)

Static completeness score — how much we know about a product. Derived from launch recency (up to 25 pts), profile richness (up to 20 pts — website, GitHub, Twitter, description), category heat (up to 15 pts), and data confidence (up to 10 pts). Always present; never NULL.

Market Signal (0–100, or NULL)

A weighted basket of independent market-evidence signals. NULL when fewer than 2 signals are present — products with too little data get no market score rather than a misleading number.

Composite (signal_score_v2)

When Market Signal is available: 0.3 × Maturity + 0.7× Market. When Market Signal is NULL (insufficient data), the composite falls back to Profile Maturity only — the product's completeness is its score until enough market evidence accumulates.

The legacy signal_score column (the live-displayed score) uses an older single-index formula. The two-index model runs in shadow mode until a backtest confirms it outperforms the legacy score; then we flip.

2. Normalize Over Present Signals

An absent signal — one where we have no data at all — is excluded from the denominator entirely. It does not contribute a zero that would drag down the average. A "measured zero" is different: if GitHub tracking is active and star growth is flat, that isa real zero and it counts. The distinction prevents a product from being penalized simply because a data source doesn't cover it yet, while still counting genuine evidence of stagnation.

3. Signal Weights

Signals are ordered by how difficult they are to fake. Higher weight = harder to manufacture. The "full credit" threshold is the raw count at which a signal earns its maximum normalized value (1.0); above it, the score is capped.

Signal	Weight	Full-credit threshold	Why
funding	1.5	—	SEC EDGAR-verified or click-testable source URL required — hardest signal to fabricate
adoption	1.4	500	G2 / Capterra review counts + rating from real-user platforms
github	1.3	1000	7-day star delta; requires ≥ 2 snapshots — measured flat is a real zero
externalRank	1.2	—	Consensus percentile across multiple third-party ranking lists
hiring	1.1	20	Net new job postings in 7 days from ATS snapshots
ship	1.0	8	Changelog / release posts in the last 30 days
press	0.9	10	Unique press-mention URLs in the last 30 days, domain-matched
hn	0.7	50	Hacker News mentions, domain-matched (not name-matched), last 30 days
bluesky	0.4	50	Bluesky mentions, domain-matched — weighted between HN and Reddit
reddit	0.3	50	Reddit mentions, domain-matched — demoted; easily gamed by founders

4. Sufficiency Gating

A Market Signal score is only computed when at least 2signals are present. Products below this threshold show "not enough info" on the market dimension — never a fabricated number from thin evidence. This is intentional: a score built on one low-weight signal (e.g. one Reddit mention) would be noise, not signal.

minPresentSignals = 2

Market Signal score is NULL when present signal count < 2. Composite falls back to Profile Maturity (always non-null).

5. Sources & Cadence

All mentions are matched by product domain, not name. Name-matching is trivially gamed by founders; domain-matching requires the mention to link to or name the actual product URL.

Source	Cadence	Notes
Product Hunt	Daily	Launch scores, upvotes, maker activity
Hacker News	Daily	Mentions domain-matched to product URL
Bluesky	Daily	Mentions domain-matched; anti-gaming same as HN
Reddit	Daily	Mentions domain-matched; weight demoted (gameable)
Press	Daily	Tech news crawl — unique URLs per product
GitHub	Daily	7-day star delta; requires ≥ 2 snapshots in window
SEC EDGAR funding	On ingest	Verified by CIK or click-testable source URL
G2 / Capterra reviews	Weekly	Top-4,000 cohort; review count + rating
Liveness probes	Daily	HTTP status + DNS check; domain-matched

6. Known Limits

—Coverage varies by signal. GitHub tracking requires a known repo URL; G2/Capterra only covers the top-4,000 cohort; social mentions depend on domain availability (~12% of products have a verified website URL today).
—The two-index model (signal_score_v2) runs in shadow mode. The live-displayed score is the legacy single-index formula. We will flip when a backtest validates the two-index model outperforms it.
—Category heat is editorial (values range 2–15 across 15 categories, e.g. ai-tools = 15, hardware = 2). It affects the legacy signal_score (via the assembleSignalScore formula), the legacy velocity_score, and the v2 Profile Maturity score. It does NOT feed the Market Signal basket — market evidence is independent of category.
—Funding scores require provenance. Rounds without a SEC CIK or click-testable source URL are excluded from the market signal — this means some real rounds are conservatively excluded until provenance is established.
—Scores are algorithmic estimates from public data. They are not financial advice and do not reflect private revenue, internal roadmaps, or non-public funding.