Attention, explained.

Windows and normalization

Choose 1h, 6h or 24h. Eight complete blocks end at the snapshot time. The median of the first four is the baseline; the last four measure persistence. Growth compares the final block with that baseline using ln(1 + current) − ln(1 + baseline).

Positive growth is converted to a percentile within the same chain, window and age band (6–48h, 48–168h, or 168h and older). Ties receive their middle rank. At least five eligible peer identities are required in this fixture implementation. Flat or declining growth maps to zero. Filtering the feed never changes the cohort.

Liquidity uses a logarithmic scale from USD 5,000 to USD 100,000 and saturates there. It remains a proxy, not a measurement of executable depth. Emergence declines linearly from 100 at six hours to zero at 720 hours.

Persistence requires breadth

A recent block qualifies only when volume, swaps and distinct active addresses each exceed 1.2 times their baseline (with a baseline floor of one), with at least 20 swaps and 10 addresses. Persistence is qualifying blocks divided by four. One volume-only spike does not satisfy this condition.

When a token is not ranked

Minimum eligibility: a valid Base address, six hours of sourced age, USD 5,000 liquidity, 20 swaps, 10 distinct addresses and complete required history. Volume, transactions, participation, liquidity and persistence must all be usable, and data confidence must reach 80%.

Observations older than 60 minutes, from the future, invalid, missing or conflicting cannot supply a score component. Conflicting required evidence withholds ranking. Missing holder growth is excluded and weights are renormalized; it never becomes zero.

Confidence is the sum of available, fresh, non-conflicting component weights, multiplied by 0.8 once if any scoring input conflicts. Both conflicting values remain visible. Unranked tokens retain their reports and reasons.

Context thresholds

Review flags identify liquidity below USD 10,000, liquidity declines of at least 50%, selected-window volume/liquidity ratios of at least 10, top-holder shares of at least 50%, or largest participant shares of at least 40%. Other flags cover contract controls, unverified source, paid promotion, stale/conflicting data and insufficient history.

What this preview does not establish

Fixtures demonstrate calculations and failure handling. They do not validate predictive performance, identify distinct people, or establish manipulation. Transaction-loop detection, social signals, narrative classification and live data ingestion remain unavailable.

The landing page preserves the earlier six-component attention-v1 illustration. The discovery feed and token reports use only the experimental v2 model described here.