Formula
5000 is the calibration target — a raw contribution sum of 5,000 normalizes to 10.0. Real-world rug setups often reach raw sums of 20,000+; they all clip to 10.0.
Level bands
Thelevel field bands the numeric score into four categories for UI:
Status
A
partial_data score is a true lower bound — every signal in missing_signals[] is one that could have added to the score had its input been available.
A failed read is not missing data
missing_signals[] holds two different things: signals whose input does not exist yet, and signals whose read failed (a database error or timeout). The second kind is also listed in unavailable_signals[], which is always present:
The risk block embedded in
/tokens/{mint}/intel carries the same status, missing_signals and unavailable_signals.
The seven holder-derived signals (single_holder_50pct, top10_high, top10_very_high, snipers_pct_high, insiders_pct_high, dev_held_high, dev_held_very_high) need a complete holder census of the mint. Until the mint has one, they sit in missing_signals[] rather than being evaluated over a partial holder set. That happens when the mint has never been censused, when the last census missed holders, or when it is too large to scan. A partial set would read a holder it never saw as holding 0%. If the check of the census itself fails, the seven are listed in unavailable_signals[] instead: retry.
Signals
Twelve signals across six categories. “Graded” signals scale their contribution with the magnitude of the underlying value (e.g.dev_held_high contributes more when the creator holds 30% vs 6%).
Holder concentration (graded)
LP + authority (boolean)
Sniper concentration (graded)
Insider concentration (graded)
Creator behavior (graded)
Metadata (boolean)
Total maximum raw contribution: 53,500 (every signal triggered at full strength).
Why this design
Glass-box so you can audit our judgment. If you disagree with a weight or threshold, you can subtract that signal’scontribution from the raw score and renormalize. The value field on each signal is the concrete measurement, not an opaque score — verifiable against the underlying on-chain data.
Stacking thresholds (e.g. top10_high + top10_very_high) compound penalties for severe concentration without making any single threshold a cliff.
Graded contribution keeps small differences in the underlying data from causing large score jumps. A token at 51% top-10 concentration doesn’t suddenly leap past one at 49%.
Phase 0 strict precedence for dev_held_*: we only fire these when tokens.creator_source is populated. We never fall back to “largest holder is probably the creator” — that produced false positives in pre-Phase-0 testing on tokens where the largest holder was a CEX hot wallet or the AMM pool itself.
Re-computing the score
The response is glass-box on purpose. If you want to weight signals differently:signals[*].value, ignore the weight/contribution, and feed the raw values into whatever you’d prefer.
See also
GET /tokens/{mint}/risk— the dedicated risk endpointGET /tokens/{mint}/intel— slim risk summary embedded in the dashboard payload- Quotas & errors —
/riskis Free-tier rate-limited; no per-endpoint cap