Convexly vs OrcaLayer vs Airavat vs Polysights vs Predicts.guru: who publishes the rule behind the label

Five readings of one public record

None of these five products has data another one cannot get. Polymarket's fills settle on Polygon, the record is public, and all five read it — directly, in OrcaLayer's case, or through the venue's own APIs. So a comparison of coverage is a comparison of nothing: it is the same chain, the same venue, and on all five cards cross_venue is false.

What differs is what each one does on top of that record and how much of that it will tell you. Four of the five rank or score other people's addresses, and one of those calls the output a finding about information. That claim is the purchase, and the grid above cannot price it. What follows is the three things that decide between them and appear on no single card.

Their own counters disagree by a factor of thirteen, except the one that matters least

Four of these five print a corpus counter on the landing page, and read side by side they are a useful accident.

The wallet populations agree. OrcaLayer puts 3.1 million addresses with at least one fill, Predicts.guru about 3.2 million wallets, Polysights 3,210,363 unique traders — inside four percent of each other. The fill counts are close enough to be the same object counted with different cut-offs: 1.4 billion at Airavat, over 1.5 billion at OrcaLayer, 1.88 billion at Predicts.guru.

The market counts are not the same object at all. Polysights says 160,722 markets. OrcaLayer says roughly 1.6 million. Airavat says over 2.2 million. That is a factor of thirteen between the lowest and the highest, over one venue's history, and none of the three publishes what it is counting — a question, a binary outcome pair, a leg of a multi-outcome set. Anything you compute per market, including a win rate, is computed over a denominator these products do not define the same way.

The sharp version of that: the one unit all four agree on is the address, and the address is precisely the unit that does not correspond to a person.

"Published methodology" means four different things here

All five would say they are transparent. Held to what a reader can actually do with what is published, they fall into four grades, and the grades are not close.

Rules you can apply yourself. OrcaLayer states its exclusions as numbers: a farmer is an address whose average buy price sits above 95 cents, a bot is one with more than 500,000 trades at an average size under 20 dollars, and its Smart Money badge requires all three of a per-market win rate at or above 55 percent, positive total profit and loss, and neither classification. Win rates are NegRisk-corrected, profit and loss is FIFO. The code and the model weights are not published, so this is checkable in outline rather than reproducible line by line — but a reader can say what the badge means and can argue with the thresholds.

Rules plus the results of testing them. Convexly publishes the methods index — methodology papers, registered forward tests, and, the part nobody else here has, published negative results. One of those negatives is about its own instrument: a pre-registered forward test of its Market Trust rating returned no separation, and the rating surfaces were frozen on 5 September 2026 rather than quietly re-tuned. Its Edge Score is a composite of three named pillars scored against a frozen 8,656-wallet comparison set. The limit is stated on its own card and matters: the Polymarket reference cohort behind the headline scores is not published, so the Polymarket results are recomputable by Convexly and not by you. Only the Manifold half ships a reproducible bundle.

Inputs named, composition not. Airavat scores every trader 0 to 100 from realised profit and loss, dollar-weighted win rate, Sharpe and Sortino, profit factor, maximum drawdown and Kelly. That is an itemised list of ingredients with no recipe: nothing published says how seven metrics become one number, what the number was fitted on, or what happens to it when a wallet has twelve resolved positions instead of twelve hundred. It is also the only product of the five that emits a directional view on a specific contract, which is a different kind of output from a record.

Nothing. Polysights publishes no methodology, no false-positive rate and no worked example for the screen it calls Insider Finder, which is also the marketing centre of the product. It publishes no terms of service and no privacy policy either — both links land on an "Under Construction" page — which is unusual for any product and pointed for one whose own FAQ says it handles the wallet on its side. Predicts.guru belongs in this grade too, but honestly: it publishes no scoring model because it has none, and its own notes say no single metric tells the full story.

The inversion is the reason this page exists. The strongest claim in the category is made by the product that publishes the least about how it is computed, and the most cautious claim is made by the product that publishes its own failures.

What each product computes, and what does not follow from it

Every one of these labels is a function over settled transfers. It observes that an address bought an outcome token at a price, before or after some other transaction, and that the token was worth something at resolution. It does not observe who holds the key, how many keys one person holds, whether the address is a fund, a shared account, a bot or a router, or what anybody knew. A page that says an account was informed has crossed from what the chain records to what a person knew, and nothing in the data covers that step. The concepts are laid out in what wallet tracking shows.

Two of the products have already demonstrated the gap, on themselves, in public.

OrcaLayer removed four Polymarket protocol router contracts from its rankings platform-wide after one of them was surfaced as an apparent insider. A contract that routes other people's orders satisfied the pattern, because the pattern is about transaction timing and a router is where other people's transactions pass. Nothing in the label could tell the two apart; a human noticing did.

Convexly published a July 2026 note that points its skill-versus-luck filter at its own top-50 leaderboard and reports that zero of the 35 readable wallets clear its own bar. The same company publishes both surfaces, and says in as many words that a historical diagnostic does not establish future profit.

Read Polysights against those two. Its own card records that the news coverage it cites is about the phenomenon of tracking Polymarket wallets, not about whether this implementation identifies anyone correctly. From outside there is no method, no error rate and no example — only the screen and the word on it.

What counts as one wallet

This is where the rankings actually diverge, and the five positions are genuinely different.

One removes. OrcaLayer's April 2026 audit of 470 million trades found that six of ten wallets with apparently flawless records were airdrop farmers — accounts buying near-certain outcomes at high volume to harvest a cent, which produces a near-perfect win rate with very little exposure. That is the top of an unfiltered leaderboard being majority artefact, by the vendor's own count, and it is the strongest single argument for paying anything at all in this category.

One bounds. Convexly does not de-duplicate people either, but it publishes the completeness of every read instead of implying there is nothing missing: its board of 12 September 2026 had 31 of its 42 wallets read in full, and every row says whether it came from a full on-chain read or from a recent-trade window that omits already-redeemed positions. Its census is explicit that a further 100 full reads did not finish and are excluded. You get a denominator rather than an identity.

Two disclose and do nothing. Predicts.guru states it plainly — the ranking is by address, one trader running three addresses appears three times, a shared account or a bot appears once. Airavat's card says a trader who hedges elsewhere or runs a second address will be scored on a fraction of their book and the score will not say so. Both are honest and neither is fixed.

One says nothing either way. Polysights publishes no position on the question.

Note what is not on this list: nobody links addresses to people, and nobody claims to. The choice is between a ranking with a published removal rule, a ranking with published error bars, and a ranking with a warning on it.

What you can examine before paying, and what the price is per

The gate runs the opposite way to the evidence, which is worth knowing before the free tiers are compared on features.

Predicts.guru is open to anyone with a browser: no account, no tier, no metering — and its terms cap total operator liability at zero dollars, which is the honest price of that. OrcaLayer gives five lookups to a guest and twenty a day to a free account, both without a card. Convexly's first check needs no signup at all. Polysights opens three of five screens to nobody in particular, and keeps the two that would make it a trading client behind a code whose price is published nowhere. Airavat opens nothing: the site is a request-access form, the terms say accounts are invite-only, and the FAQ entry that asks what it costs does not render its answer to a logged-out reader.

So on the one axis where a reader can protect themselves — look at the output before paying for it — the two products with the least published method are the two that show the least.

The paid units are also not the same unit, and the two nine-dollar prices are the trap. OrcaLayer's 9.99 a month removes a cap on how often you may look, and adds Telegram alerts, a 25-wallet watchlist and CSV. Convexly's 9 a month buys three wallets that it re-reads in full every 30 days; its 49 tier buys ten on the same cadence plus on-chain reconciliation and an API; its 499 tier buys fifty re-read weekly. One is priced per query, the other per maintained record, and the second is the one that decides whether a number you look at in November describes November.

Convexly's 19 USD one-time read is the outlier worth naming: a single reconciled read of one named address delivered with the block it was reconciled at, a hash over that subset and the command that re-derives it against any Polygon archive node. It is the only artefact any of these five sells that a buyer can independently check.

Where the cards use the same field differently

Two gated products, two opposite flag conventions. The Polysights card sets charting, order book, live trading and news to false and explains in prose that all four are advertised and all four sit behind the access code — a flag is what a reader filters on, and a reader filtering for live trading wants something they can trade on today. The Airavat card sets charting and order book to true for surfaces that are equally unreachable without an invitation. Filter the category on either flag and you will get one of these two and not the other, for the same underlying situation.

The verification dates line up, unusually: all five cards were re-read between 19 and 20 September 2026, so nothing here rests on a stale check.

Which one to take

Take OrcaLayer if you are going to act on a ranking. It is the only one of the five that both publishes its exclusion thresholds as numbers and has published what the unfiltered population looks like — six in ten flawless records being farmers is the fact that makes every unfiltered leaderboard in this category unusable for the purpose people use it for. Twenty free lookups a day is enough to decide, 9.99 removes the cap and adds alerts, 19.99 is the API tier.

Switch to Convexly the moment the question changes from "who is up" to "is this record distinguishable from luck". Those are different questions and OrcaLayer answers only the first. Start with the 19 USD single read rather than a subscription: it is one wallet, it comes with the block and the command to re-derive it, and it is refundable within 14 days. Move to 49 a month only when you need ten addresses kept current rather than one answered once. It costs more per wallet than everything else here and the reason is on the card — the price is the re-reading, not the watching.

Take Predicts.guru if you will supply the judgement yourself. Free, no account, the full per-address history and ten-way comparison, and its own notes tell you a row is an address. There is no API and no export, so the ceiling is what you can read off a page, and its header carries a persistent link to the venue it covers — the data is not wrong, the framing has a direction.

Airavat is not a purchase you can currently make, and if you do get an invitation the reason to use it is not its trader score, which is the weakest-documented of the four that document anything. It is the UMA layer: active disputes, the resolution lifecycle and diffs when a market's rules text changes, over a stated 3,700-plus disputes. A quietly reworded resolution criterion is the failure mode that costs a reader a position they thought they understood, and nothing else in this category watches for it.

Polysights is worse on every axis this page measures — no published method, no error rate, no terms of service, no privacy policy, no API, no export, and two of five screens behind a code with no published price, on a landing page that carries "V1 Coming Soon" at the top and a sentence saying everything above is live in the terminal right now at the bottom. The single axis on which it is not worse is that it is the only one of these five that intends to put execution beside the wallet screen at all: Convexly deliberately will not trade, OrcaLayer does not, Airavat simulates fills against virtual funds. That intention is currently a code form and a contradiction on one page. If execution next to the analytics is the actual requirement, Polyrama does it today and reads two venues while doing it.