Every pick tracker I have used asks the same question: which tipster can I trust? None of them asks the one that costs me money: can I trust myself?

So for the Total Recall hackathon (Tatum x Walrus) I built Priors, a forecasting coach that remembers every pick I make, why I made it and how it ended, then argues with me before my next one.

Live demo: https://usepriors.vercel.app
Code: https://github.com/Jennivarl/priors

What it does

You log a pick before the event: what you back, how confident you are, the price, and why, in your own words. Before you commit, Priors gives you a second opinion from your own record:

Your gut says 80%. You overrate Arsenal by 41 points over 5 picks. On football you run hot.
Coach: 24%. Market: 51%. Corrected for your bias, the market is more right than you.

Over time it builds a bias fingerprint: overall calibration (on the demo record I say 71% and 50% happens), bias per team and per category, tilt after losses, and how I do when I disagree with the market by 10 points or more. Small samples are shrunk toward zero, so two lucky picks do not become a bias.

How it uses Walrus Memory

Walrus Memory is not a database here. It is the part that makes the learning possible.

  • Every pick, reason, result and lesson is a memory on Walrus Mainnet. The lessons the coach learns are memories too, so you can see what it learned and when. The demo record holds 129 memories.
  • It learns from how you think, not just your numbers. Reasons live in their own namespace. When you type a new reason, Walrus semantic recall finds past picks where you argued the same way, even in different words ("great form, they're due" matches "they're flying, they owe us one"), and shows how that thinking worked out. I calibrated this on real recall distances: the same line of thinking lands at 0.41 to 0.55, unrelated reasoning at 0.65 and above, so the echo counts matches below 0.6.
  • Walrus Analyze extracts facts from the reasoning text.
  • It is portable. One button wipes the app's local data and rebuilds everything from Walrus alone. In testing it restored 25 picks, 24 results and 7 lessons exactly.

How it uses Tatum

  • Proof of time. Every pick is sealed with the live Sui Mainnet checkpoint height and digest, read through Tatum's Sui gRPC gateway, plus a sha256 commitment over the pick's content. A digest cannot be known before its checkpoint exists, so the pick was made after it, and the Walrus write shows it existed by then. Editing any field later changes the commitment.
  • Onchain prices. Crypto calls on SUI and WAL ("SUI above $0.836 in 3 minutes") read Cetus pool objects on Sui through Tatum, both when the call is made and when it settles.
  • I used Sui gRPC, Tatum's recommended path now that Sui is moving off JSON-RPC, with a rate limiter for the free plan's 3 requests per second.

Results nobody can fake

  • Crypto calls settle themselves at the deadline from onchain pool state. The settlement keeps the price and the checkpoint as evidence.
  • Sports and politics picks use live Polymarket prices. The server re-quotes the price itself, so nobody can type a fake one, and the pick settles when the market resolves.
  • Manual picks still exist for anything not listed, and they are labelled self-reported everywhere.

In a live test I placed 8 real crypto calls with two styles of reasoning. Data-driven reasons ("low volatility, a 0.2% move in 3 minutes is unlikely") won 5 of 5. Hype reasons ("it's due, everyone is bullish") won 1 of 4. Nobody chose those results; the chain did.

How it learned over time

The coach is a small, tested set of pure functions. Each time a result settles it recomputes calibration, bias per team and category, tilt after losses and performance against the market, and writes any new or changed lesson to Walrus as a new memory. The second opinion starts from your gut and subtracts what your record says you get wrong, then adds the reasoning echo from Walrus recall.

You can watch it happen on the real calls: after the hype reasons lost, typing "SUI is going to pump, it is due" brings back those losses from Walrus and pulls a 75% call down to 60%.

Honest limits

  • The live demo record mixes a seeded demo history (events marked "(demo)", illustrative results, but every pick still really sealed on Sui and written to Walrus) with real calls that settled onchain.
  • Polymarket is not on Sui. It is the deepest public market for sports prices today, and a Sui-native prediction market can plug into the same result-source interface.
  • A pick is time-bounded by a checkpoint digest and a Walrus write, not anchored in its own Sui transaction, which would need a funded wallet per user.

What is next

Results come from three sources behind one interface (onchain price, market resolution, self-reported), so a Sui prediction market is the next source. After that, a weekly bias report, and a safer-gambling module for operators: the tilt alarm is exactly the kind of signal they are expected to provide.

Why "Priors"? In statistics your priors are the beliefs you walk in with. Good forecasters update them with evidence. Priors keeps the evidence, so you actually do.

Built with Walrus Memory and Tatum for the Total Recall hackathon.