Scoring changes are marked where they altered stored results, so you always know whether a report you are looking at predates one.
Every scoring weight, the calibration percentiles, and the measurements that came out badly are now on the site rather than in a README. Pricing is prepaid periods with an expiry rather than a subscription, because crypto has no direct debit and pretending otherwise would only surprise people later.
Scoring engagers by follower count, account age and ratio turned out to be inverted on hand-labelled data — the 'better' accounts were frequently the farmed ones. Reading what people actually wrote took the evaluation from 78% to 91% against a 61% trivial baseline. Account-level signals remain, weighted at roughly what they earned.
The engagement threshold had been guessed at 12 per thousand followers. The measured p75 is 1.4, which had made two of the four profile quadrants structurally unreachable. Thresholds are now re-derived from the distribution rather than hard-coded.
Slippage was charged to the asset as a full round trip while the peer benchmark was charged nothing, which penalised thin pools against an uncharged median. Liquidity is also reported as of now rather than as of the post, so it is scaled and bounded. Fill costs are shown but deliberately kept out of the scored return.
Absolute return made every call during a market-wide rally look like skill. Performance is now measured in excess of a peer median sampled on the same chain over the same window, which is also why a losing call in a worse week can still score respectably.
Roughly 31% of a KOL's token posts are not calls at all. Scoring them was dragging every average toward the middle, so posts are now classified before they are scored and non-calls are excluded from aggregates.
Every stored result now carries the scoring version it was computed under. Rescoring is explicit and logged rather than silent, so a report you read six months ago still means what it meant then.
Older work predates the corpus being usable for comparison and is not listed. If you need to know exactly how a specific stored result was computed, it carries its own version stamp — send it to us and we will tell you.