How the numbers are made.

01 Verified from venue data

A record on TAIL is built from venue fills, never from claims. Kalshi records read from the exchange’s own fills API after a caller connects — read access, non-custodial, revocable at any time. Polymarket records are public on-chain activity: a wallet address is enough, and no connection is required to compute one.

There is no self-reporting anywhere in the pipeline — no screenshots, no spreadsheet uploads, no manual entry. The same ingest that writes a caller’s wins writes their losses, and once a fill is ingested, TAIL cannot edit it any more than the caller can.

02 What a receipt says

A receipt prices one closed position. Entryis the volume-weighted average of the fills that opened it — what the caller actually paid per contract, not the market’s best print. Exitis the volume-weighted average of what closed it: sells at their real prices, and any contracts held to resolution at the payout — 100¢ on the winning side, 0¢ on the losing side. The return is the move between the two:

return = (exit − entry) / entry

Receipts mint on final resolution only. Polymarket long-form markets resolve through an optimistic oracle with a challenge window of about two hours; a proposed outcome is not an outcome, and nothing mints until it is final. Kalshi settlements are exchange-final. The pipeline treats proposed and final as distinct states, because a receipt that later flips would be worse than no receipt at all.

Open positions are never public. And when a caller has live copiers, each receipt goes public on a short delay after the position closes — 15 minutes by default — so a record stays auditable without becoming a front-running feed.

03 Calibration

Every entry price is a probability claim: pay 62¢ for YES and you are saying the event is at least 62% likely. Calibration scores those claims against what actually happened, using the Brier score — the mean squared gap between the price paid and the outcome:

brier = mean((p − o)²)p = entry price / 100 on the YES-equivalent side · o = 1 if it happened, 0 if not

What displays is that score rescaled, so the anchors are readable:

calibration = (0.25 − brier) / 0.25clamped to [−1, 1]

0 means the caller prices like coin flips. 1 is perfect foresight. Negative is worse than flipping. Nothing displays under 25 resolved positions — small samples flatter everyone.

04 The score

score = winsorized_mean(return_bps, 5%) × ln(1 + resolved_count) × liquidity_factorliquidity_factor = clamp(median_position_notional / $25, 0.5, 1.0)

winsorized mean

The extreme 5% of returns are clipped before averaging — one moonshot, or one wipeout, cannot manufacture a record.

ln(1 + resolved_count)

More resolved calls raise the score, with diminishing returns — a 100-call record beats a 25-call record, but not by 4×.

liquidity factor

A median position under $25 scales the score down, to half at worst — farming a record out of dust-sized positions does not pay.

Scores are computed over 30-day, 90-day, and all-time windows, overall and per category. A caller is ranked only with at least 25 resolved positions in the window and an account at least 14 days old.

05 Losses shown

Every resolved position mints a receipt. There is no deleting a bad call, no archiving it, no choosing which trades count: if it resolved, it is on the record. Losses render at the same size and the same weight as wins — a red number is never smaller than a green one anywhere on TAIL. A record that only showed wins would not be a record.

06 Integrity policy

Three patterns are flagged automatically, alongside every stats run:

parlay noise

Multi-leg and combinatorial positions display on a record but are excluded from the score by default, shown separately. Copying mirrors single binary markets only — parlays are never mirrored.

wash suspect

Repeated offsetting buys and sells in the same market within 2 minutes.

anomaly

A stake more than 10× the caller’s trailing median.

A flag means an under reviewbadge on the leaderboard — visible and stated, never a silent deletion. Removing a caller quietly would be its own kind of lie.