[Flotapponnier]

Follow-up to #2717. Florent said venues are missing from the perp open-interest table and named Ondo. The missing venues were a symptom; the cause is the headline.

The column was mislabelled, and the 49% gap is what that looks like

The header promises each venue's open interest. The column was reading bench 265's oi panel, which is DefiLlama's open-interest overview keyed by protocol. That is a mislabel, not a discrepancy, and it had two consequences:

• The values disagreed with the venues. Hyperliquid read $8.57B against the venue's own $12.77B, a 49% gap on the largest row on the table. EdgeX $225M vs $452M, Pacifica $57M vs $112M, Extended $92M vs $185M, all roughly half; Aster $1.43B vs $1.22B, the other way. Two different measurements, not one with noise.
• The cohort is token-gated, so a venue without a ranked token had no row at all: trade.xyz $3.76B, GRVT $472M, Lighter RH $228M.

The column now reads the perp cohort harness (perp_venue_oi_usd, perp_venue_volume_24h_usd), each venue's own API, through fetchPerpCohort() — the worker-written snapshot /perps uses, so still no Prometheus call at render time. Open interest, volume and turnover now share one source. The page states whose figure it is, says the column changed on 29 September and why, and the reading lines say DeFiLlama's overview reads differently and is not what these columns show. No averaging of the two, no silent switch.

The one seam left, and it is written down in four places so nobody rediscovers it: turnover is bench 271's published ratio, which divides the 24-hour averages of those two gauges, while the columns beside it show the latest read. Same gauges, same harness, different window, so volume / OI as displayed does not equal the turnover column, and the gap is widest on a venue whose book moves during the day. Recorded in bench 271's methodology, the table's note, CAPITAL_READING.openInterest, and the comment above perpOiRows. The ratio is read from the bench rather than recomputed per page so one named ratio has one value across the site.

No 7d column on the perp table: nothing publishes a 7-day series of the venue gauge, and pairing a venue level with a DeFiLlama move would be the same seam one column over.

Truncation

.slice(0, 10) is gone from both OI tables and each heading carries its count ("Prediction market open interest, all 13"). A venue the harness gets no open interest from (Orderly, Backpack, SynFutures, Vest, KiloEx) keeps its row, ranked last, with the reason on the cell instead of dropping off a table a reader is counting.

The prediction-market 7d column is hidden by the coverage floor from #2717 (1 of 13). Twelve markers next to one value is the noise that made the page read as empty; the per-row reasons stay in /api/capital for a machine consumer, and the column returns on its own once the eleven warm-up windows fill.

Venue classification

From harnesses/perp-cohort-stats/cmd/script/registry.go, the per-venue source files, and the provider registry:

venue	verdict	evidence
trade.xyz	DEX	seeded onchain, Hyperliquid HIP-3 dex, native source srcHLXyzNative
GRVT	DEX	seeded onchain, zkSync, native source + DeFiLlama fallback
Lighter RH	DEX	seeded onchain, Robinhood on Lighter L2, native source
Ondo Perps	DEX, hybrid	off-chain matching in Intel SGX enclaves with on-chain custody on Ethereum and Arbitrum, USDC and Ondo tokenised stocks as collateral; read from api.ondoperps.xyz, already ranked on bench 041 and bench 271
Variational	DEX, not added	api.variational.io is NXDOMAIN; open interest comes from DeFiLlama alone and it publishes no volume, market count or top market

The first three appear automatically once the column moved. Ondo was never in PERP_VENUES, so /perps was dropping it too — a registry gap found through a hub audit, costing that page a $89M row. Now seeded. The "SGX" in Ondo's description is Intel SGX, not Singapore Exchange, so regulated would have been wrong; onchain matches what benches 041 and 271 already call it. Variational stays out deliberately, with the reason written next to Ondo's seed.

The Gains outlier

There is no outlier guard on the turnover column anywhere, and a comment now says not to add one, citing the sub-half-hour holding period. Bench 271 publishes it unclipped too.

Verification

fetchPerpCohort() is unreachable from a local dev env, so the perp table is absent locally and the page degrades to the prediction-market table alone (verified, no error). Production /api/md/perps confirms the snapshot and the 26-row venue cohort. The row build is extracted as a pure perpOiRows with 5 tests covering the CEX filter, the prediction-market overlap, the no-OI rows, the product-slug turnover join and the hidden 7d column.

pnpm validate 240 specs, pnpm typecheck, eslint clean, 367 tests pass.