# findings.txt: perpetual-futures-survivorship-bias (reviewed 2026-10-24)

**How this was checked.** I compared the article against its images, the author's manuscript and arithmetic. I also re-queried the public, key-less endpoints on 2026-10-24 (raw responses and retrieval times are in `requery/`; the recomputation is `recheck.py`).

From today's two Bybit calls (linear default and `status=Closed`), the 2026-09-14 snapshot rebuilds exactly:
- 762 live and 308 closed contracts, and 571 = 399 + 172.
- The 298 and 126 figures, and the cohort table.
- Lifespans of 231.0 / 440.6 / 785.7 / 1,653.7 days, all deciles, and the 127 and 225 counts.
- The live age quartiles, the 53 midnight stamps and the 2018-11-14 inverse stamp.

Other checks:
- The six archive dates in table 34 match `public.bybit.com`.
- OKX: 24 rows carry the 10:12:51 stamp across 10 dates. The FIL, COMP and BTC stamps match. The `state` parameter is ignored, and OKX's docs list only four request parameters.
- Bitget: `launchTime` is empty on every row.
- Binance: the live board rebuilds to 528 contracts, a 629.9-day median and 120 under one year.

## Checklist lines (one alt set, six locales)
- **122 en: CORRECTED.** All the numbers check out (details in verdict.json), but the wording was wrong in three places:
  - Alt 51 calls the 21:00 UTC point a "settled reading". It is a predicted rate. No venue settles at 21:00; all three settle at 00/08/16 UTC.
  - Alt 8 does not say that the 172 delisted contracts are missing from the roster its bar is labelled as.
  - Alt 23 says "50 day bins". The histogram has 33 bins of 50.1 days.

  Alt 30 now carries the caveat that the venues' listing stamps are not one measurement. The funding medians match the image and the body. The underlying 167 hours come from Athenum's own feed and could not be re-queried.
- **123 ko, 124 tr, 125 ru, 126 vi, 127 de: CORRECTED.** The numbers are verified as for 122. In these locales the alts are still in English and need localising from the reviewed alts 8, 23, 30 and 51.

## Other substantiated defects (English indices)
1. **Binance SETTLING rows misread (blocks 0, 41, 42, 44, 45, 48).** The article says Binance has "129 USDT perpetuals in a `SETTLING` wind-down and no history", described as "contracts on their way off the board right now".
   - On the re-read, 128 SETTLING USDT perpetuals carry a `deliveryDate` before 2026-09-14, going back to 2022-06-17. 86 of those dates fall inside the one-year window.
   - So these are contracts already delisted that Binance still serves: Binance does keep a partial graveyard.
   - The 129 count and the 880.3-day median fit one older row having been dropped since. That supports the sibling fact in a refined form: delisted rows partly stay and partly disappear, so counts depend on the source and the date.
   - Block 41 also measured the age of these rows to 2026-09-14, which includes the time since they were delisted (median 281 days). Measured from listing to delisting, the 128 dated rows traded a median of 498 days, and 45 of 128 (35 %) traded for more than two years. Block 41 now gives both measures, and block 44 says "only partly" instead of "not" for the delisted half.
   - Hyperliquid's `meta` endpoint also flags 56 assets as `isDelisted`. "The one large venue" is therefore scoped to "the one large venue checked here".
2. **Funding "settled" (blocks 51, 52, 53, 54).** Chart 51 has been withdrawn, so alt 51 now describes the withdrawal. Caption 52 is now a withdrawal caption that keeps the predicted 21:00 readings, the 167-hour medians and the 00:00 settlements. Block 54 no longer refers to the boxes or the chart. Its claim that each IQR is wider than the gap between medians could only be checked against the withdrawn image, and is kept as published. "Settled hour" really means a closed hourly bar. The payments that followed, at 2026-09-15 00:00 UTC, were:
   - Binance 0.003563 %
   - Bybit 0.002656 %
   - OKX 0.002675 %

   These are now quoted in the text. This is the "rate versus payment" defect. The opening sentence of block 53 is also reworded.
3. **Overlapping IQRs used as a significance argument (blocks 31, 38).** The article says "ranking the venues would be reading a difference the dispersion does not support".
   - A rank test gives z = 9.2 for Binance against Bybit and z = 10.4 for Binance against OKX. Bybit against OKX is only z = 0.9.
   - The article's own 59 % vs 23 % comparison gives z = 11.7.
   - The valid reason not to rank the venues is that their stamps are different measurements.

   Also, "still lands on the right side of a one year line" is now limited to "most contracts", because ETHUSDT's stamp is 145 days off.
4. **Basis of the flat-hazard test (block 19).** The text says it used each cohort's "own mean exposure". It actually used each cohort's mean age since listing, with expected removals of n(1 − e^(−h·age)).
   - That reproduces 4/5/13/59/208 and a chi-square of 37.8 exactly, but those expectations sum to 288, not the 308 observed.
   - On the article's own contract-days basis, the expectations are 4/5/13/53/233 and the chi-square is 54.6.
   - The 1–2 year cohort is still the only one above the flat rate (z = 6.4, which survives a Bonferroni correction across 5 cells).
   - The over-2-year cohort is below the flat rate (z = −2.3), but that does not survive the correction.

   The text now states both bases and notes that the published shortcut sums to 288. The cohort cut-offs are conventional, not picked in hindsight.
5. **Floor logic (block 13).** The article argues that the earliest delisting being dated 2022-02 "makes 172 a floor". A retention start in 2022 cannot cut into a window that opens on 2025-09-14. Where it does matter is the older cohorts, so the 55.5 % survival figure is an upper bound. The floor caveat is kept only for rows Bybit might drop without trace. The re-read returned 308 closed rows for 2026-09-14 with identical lifespans.
6. **Causal overclaim (block 20).** The article says "the contracts that drop out are the ones that stopped working", so a backtest on survivors "will report a kinder drawdown". It also says itself that the money effect is not measured. This is now limited to long-biased strategies and stated as an argument, not a measurement. The "a few quietly deleted losers" line is made conditional.
7. **"the one contract that is on every board" (block 50).** ETHUSDT and ETH-USDT-SWAP are on all three boards too, so this now reads "a contract".
8. **Stale tool count (block 59).** The article says "34 calculators"; `/tools/` currently lists 37.

**Verified and unchanged:** the title, the description, the cover chart and both tables.

**Kept as published because they cannot be re-checked:**

| Claim on 2026-09-14 | Value now |
| --- | --- |
| The 167-hour funding medians | Not re-checked in this review (Athenum feed) |
| Binance status counts 766 / 130 / 1 | 794 / 136 / 1 |
| OKX 479 rows | 522 |
| Bitget 786 rows | 857 |
| Board re-run "seven minutes apart" | Not checkable |

## Locale structure
All six locales have 60 elements with the same tags, image sources and link positions (blocks 3, 20, 27, 53, 59). No `indexmap.json` is needed.

Spec: 16 replaced blocks (0, 13, 19, 20, 31, 38, 41, 42, 44, 45, 48, 50, 52, 53, 54, 59), alts for 8, 23, 30 and 51 (51 is the withdrawal notice), and a correction note that also records the withdrawal.

Figures: chart 8 needs a new label, and alt 8 is now neutral either way. The data series behind chart 51 cannot be rebuilt (Athenum feed, no public history of hourly predicted rates), so it is being withdrawn and replaced by a neutral "chart withdrawn" notice image. figures.txt lists everything that is exactly known.