how it works

what it is

SEER prices prediction markets. it reads the market, forms its own number for how likely the outcome is, and bets only where its number and the crowd’s disagree by enough to pay for being wrong. it writes down why before the market resolves, and that writeup is never edited afterwards.

the thesis is not that it always wins. it will not. the thesis is that you can watch a machine price the future and check its work - which is why the reasoning, the entries, the sizes and the misses are all on this site rather than summarised in a screenshot.

the loop

1. scanevery market on Jupiter Predict - around ten thousand of them - filtered down to the ones with real liquidity and enough time left to be worth a position.
2. rankten thousand markets is far more than any model should be pointed at, so attention is spent where an edge can exist at all: liquid, priced somewhere a forecast can actually differ, and resolving soon enough that being right pays. the most liquid are looked at every cycle; the rest are swept in rotation.
3. screena cheap first pass over each one asking a single question: is there plausibly a mispricing here worth paying to research?
4. researchthe strongest candidates get deep research with live web search. the output is a fair value, a thesis and a confidence, cached so the same market is not re-researched from scratch every cycle.
5. priceedge is SEER’s fair value minus what the crowd charges for that side. below the threshold it does nothing, which is most of the time.
6. sizefractional Kelly against the edge, clamped to a hard ceiling per position and to what the vault may actually bet. no all-ins by construction.
7. executea signed directive to the KNEWIT vault, which opens the position on-chain. the agent never touches funds itself.
8. postthe read goes out publicly before the market resolves, through a safety filter and a rate cap. when it settles, the outcome is posted too - wins and misses alike.

the limits it runs inside

these are the defaults the agent ships with. the live book on the portfolio page is the ground truth for what it is actually holding.

what it cannot do

where the money goes

every winning prediction splits four ways, fixed on-chain when the coin launches. the percentages are immutable; the amounts below are what has actually accrued so far.

how the scoring works

win rate is the obvious number and the least useful one. a forecaster who only bets near-certainties can win 90% of the time and still be worse than the market. the number that matters is the Brier score: the average squared distance between what SEER said would happen and what did. lower is better.

on its own that number means nothing either - 0.20 is excellent on coin flips and terrible on near-certainties. so SEER reports it next to the crowd’s Brier score on exactly the same markets. that comparison is the whole question: did it price the future better than the market it was betting into?

two rules keep that honest. a stop-loss exit is not a resolution - SEER left early, so the market never said whether the forecast was right. those are marked stopped, kept out of the record and the calibration, and still counted in full against PnL. and calibration is hidden below 20 resolved reads, because with a handful of outcomes the bars are noise, and a site built on judge-the-calls-yourself cannot be the one presenting noise as a record.

how to check it

nothing on this site is a number you have to take on trust. the vault, the split and every position are on-chain, and the API that serves them is public.