Why does every Convexly read carry a disclaimer?

Because the disclaimer is part of the measurement. The API ships a fixed line with every verdict: diagnostics on a public on-chain record, not investment advice, not a forecast. It cannot be disabled, and the badge carries the same caveat in every view. A number stripped of its limits stops being true, so the limits travel with the number by construction.

What is the difference between in-sample and forward?

In-sample means measured on the record that already exists: the wallet's resolved past positions. Forward means tested on data that did not exist when the method was frozen. Every skill read on this site is in-sample and labeled that way. The forward questions (does the cleared set keep outperforming?) are answered by tests filed in advance, whose verdicts compute mechanically from time and outcomes, and those verdicts are pending while the windows mature.

What we never say

Q: Has Convexly ever published a test that failed?

Yes, and it stays published. The follow-up test of whether the Edge Score composite predicts a wallet's future rank was committed to in advance with a +0.30 bar. The frozen-coefficient leg came in at +0.111 (95% CI 0.05 to 0.18) and the refit leg at -0.082, below zero. Both legs missed the bar, the result is on the negative-results page, and the composite is presented as a descriptive ranker of past behavior because of it.

The answer first

Everything on this site describes resolved past behavior on a public record. Nothing on it is investment advice, a forecast, a rating of a person, or a promise about the future. Those are not legal boilerplate lines; they are the boundary of what the evidence supports, and the site is engineered so the boundary travels with every number.

The intuition, with one worked example

Here is the same finding written two ways. The hype version: "Our score predicts which traders win next." The receipt version: we committed in advance to a test of exactly that claim, with a +0.30 bar filed before any analysis ran. The frozen-coefficient leg came in at +0.111 (95% CI 0.05 to 0.18) and the refit leg at -0.082, below zero. Both legs missed the bar. So the public claim is the second version: the composite ranks past behavior across wallets and did not clear its own forward test, which is published in full on the negative-results page and stays published. A site that only ever reports wins is indistinguishable from a site that buries losses. Publishing the miss is what makes the other numbers worth reading.

The actual method

The discipline, as enforceable rules:

Not advice, not a forecast. Every API verdict ships a fixed line that cannot be disabled: "Diagnostics on a public on-chain record, not investment advice, not a forecast." The badge carries its caveat in every view.
In-sample says in-sample. A skill read is a property of the resolved past record. The skilled state's own definition ends with "retrospective and in-sample."
Forward claims wait for forward tests. The persistence questions are covered by tests filed in advance with an independent registry; their verdicts compute mechanically from time and outcomes, and they are pending while the windows mature. A null verdict would be published as a valid outcome.
No estimate without its interval. Enforced in the schema and the export code, not by editorial habit (lesson 3).
A word gate on every build. An automated check scans public copy and blocks a list of certainty words unless the receipt exists. The failed-test archive stays linked in the site footer.

Where you see this on the site

The footer of every page links Failed Tests next to the research itself. Method receipts and the hash-chained evidence log live at /research/preregistrations and /evidence, and /independence documents that Convexly takes no venue revenue, so no verdict here is paid for by the thing it measures.

What this does NOT mean

Restraint does not mean nothing works. It means the numbers you do see survived the checks in these twelve lessons: a price-netted measure, an interval, a structural screen, a multiplicity correction, and a placebo baseline, with the misses published alongside the hits. That is the whole idea of the site in one sentence: the confidence you can place in a number should come from the process that produced it, not from the volume at which it is stated.

Put it to work: check a Polymarket wallet

Convexly publishes new methodology research roughly every 6-8 weeks plus the /learn series on a rolling cadence. Get the next paper in your inbox when it ships: