[Flotapponnier]

Two things: the brand assets Florent supplied, and Nova wired into the perp cohort.

The assets

public/logos/nova-markets.jpeg and an updated public/logos/route.png, in the logos directory rather than at the repo root where they arrived. The SVG I had extracted from Nova's site is deleted, and nova-markets comes out of NEEDS_DARK_CHIP since the supplied image carries its own ground.

Nova in the cohort

Nova runs the same gTrade software as Gains on Arbitrum: the same two backend endpoints, the same payload shapes, the same 493-pair registry. So it is read by the Gains source pointed at its own hosts rather than by a 265-line copy that would need keeping in step. It appears on no DefiLlama page of any kind (checked across all 8795 protocols on 2026-10-22), so the native backend is the only path for every figure, and it carries no TVL or 30-day volume at all.

Open interest is new for both venues. Each holds it in collaterals[].pairOis[] and neither source read it. Gains keeps DefiLlama above the native read in the router so its published row does not move; Nova is on no DefiLlama page, so the native read is the only one there. Each collateral converts at its own decimals and USD price, and the read refuses rather than returning a zero when a collateral declares neither, because a zero reads as "no positions" rather than "not measured".

Where it lands

Bench	Treatment	Nova reads
041 volume share	ranked	$1.38M, last, against aevo's $2.13M
116 active markets	measured, unranked	493
271 volume / OI	excluded, reason in the methodology	would be 123x

116 is the one worth reading. Nova and Gains ship the identical 493-pair registry. gTrade delists a pair by setting its max leverage override to 0.001x rather than removing it: Gains carries that override on 315 pairs and counts 178, Nova carries it on none, and 40 of Nova's pairs have no override entry at all. Nova's 493 is therefore the catalog it inherited, not a book anyone has confirmed is openable, so ranking the two together would publish "Nova lists 2.8x the markets of Gains" off one list. The row ships with unranked: Registry unpruned, which keeps the measured value visible and out of the ranking.

The product page's Active markets card computes its own rank from the cohort, so it could have disagreed with the bench silently. It now shares the exclusion, which is the same shape as the CEX filter already sitting two lines above it, added when Binance's pair count outranked every DEX on a card linking to a bench that lists no CEX.

271 is excluded for the opposite problem to the usual one: both series exist and the denominator is too small to divide by. $11,269 of open interest across a single pair against $1.38M of 24h volume is a ratio of 123x that would top the board, above Papertrade. That describes a venue with almost no open positions rather than one turning its book over 123 times a day, and one position opening would halve it. The methodology records why, next to the existing bullet listing the venues excluded for publishing no OI at all.

Also maps Nova's hype-1000 group (one pair, HYPE1000/USD) to crypto. Without it one unknown group withholds Nova's whole breadth read, which is the right default and the wrong outcome here.

Cache keys move in lockstep (bench-unfiltered-v93 / all-benchmarks-v86): a cached entry holds the provider set current when it was written, so a bench gaining a row needs the key to move or the row never appears. bench-filters stays at v35, since none of the three benches declares dimensions.

Two corrections to earlier readings in this work

• I had Nova's active markets at 1, from a field counting pairs with open interest. Bench 116 counts the max-leverage filter instead, which is 493. The difference is the whole reason the row is unranked rather than simply last.
• I had Nova's TVL at $119,650 from DefiLlama nova-nlp. That is "Nova NLP", a risk curator on Hyperliquid L1, an unrelated protocol matched on the name. Nova has no DefiLlama entry, so it is absent from bench 144 because there is no TVL source at all, not because of that bench's $1M floor.

Checks

pnpm validate (253 specs), typecheck, lint (0 errors), test 684 pass, and go build / go vet / go test green in the harness.

Six new tests, each verified against a mutation of the code it covers:

• the OI conversion sums every collateral at its own price, and refuses a payload missing decimals or a price, while still reporting a real zero as measured
• the two gTrade configurations cannot share a venue, a source label or a host, which is the quiet way to publish one venue's numbers twice under two names
• the unranked card neither ranks Nova nor pushes Gains down a place, and the CEX exclusion survived the new filter
• the exclusion set and bench 116's spec cannot drift apart, since both encode the same judgement in two places

The live test against both real backends is behind OCB_LIVE=1, following the leverage-terms precedent, and prints what each venue publishes so a spec can be written against it.

Still to do after merge

perp-cohort-stats needs redeploying on the VPS and the materialize worker rebuilding before any of these rows carry a number.