Portfolio • Case Study 02
Glimps
What is it?
A local-only crypto teaching tool that trades. Each morning it gathers vetted evidence, explains how and why money is moving, decides behind hard rules in code, and measures every decision against simply holding.
Problem framing
I’m a retail investor, not a trader, and I wanted to understand the market rather than be told about it. Most tools hand you a conclusion with no way back to the evidence, and the better they sound, the harder they are to check. Here, a wrong input costs real money.
So Glimps is an information-quality product first. If the information is corrupt, everything built on it is corrupt; there is no slightly wrong. A twelve-principle charter holds that line, and it wins over speed and convenience every time.
And the question comes before the money. Whether the system has an edge can be answered without a single live order, so none is placed until it is.
Research as evidence
Research isn’t a phase here; it’s what the system runs on. Evidence comes only from a vetted list of free public sources, and each reading is stored with its source and the time the event happened, so lag is visible. No data beats unattributed data. The evidence can’t be edited or deleted, the model never writes to it, and every order has to cite a reading by its id.
The harder part was keeping the system honest about itself:
- A fair benchmark. The account is measured against just hold: the same money doing nothing but investing deposits by a fixed rule. An earlier alpha number was retired when I saw it flattering the system.
- Every decision scored. Each order is scored at seven, thirty and ninety days against not making it, and so is money left waiting, so holding back has a cost on the record.
- Screen versus model. Code ranks every coin on nine signals before the model sees anything, and that ranking is scored too, so I can tell which one adds the value.
- The real account, not a copy. I kept trading by hand, and the simulator was working from a copy made on day one, so the brief was planning around cash I’d already spent. Now every run reads the exchange first and applies my trades to both sides of the comparison, so they count for neither.
- Earned confidence. Every prediction carries a probability that is scored when its horizon passes, and each brief says what the day’s evidence doesn’t cover.
Solution as a system
Glimps is a daily loop with a check at every joint. Evidence is gathered and frozen, code ranks the coins, the model writes the brief and decides buy, sell or hold, and hard rules in code have to pass before anything fills in a simulator that follows the real Kraken account, hand trades included.
Security is part of the design. Custody stays at the exchange, with withdrawals disabled on every key at the exchange itself, not in my code. The server answers only on the machine it runs on and asks for a sign-in, secrets never enter version control, and the runtime depends on five pinned packages.
Live money is a gate, not a date. The switch stays locked until the simulation beats just hold three months in a row and beats at least one falling month, and even then I flip it by hand. Once live, trading halts on its own if the account trails just hold too far in a month, and any order with an unknown outcome stops everything until it’s resolved.
Takeaway
Information quality beats speed. The slower, auditable pipeline makes decisions I can defend and, more importantly, prove wrong. Without provenance, it’s just vibes.
The edge isn’t demonstrated yet. The benchmark is hard to beat on purpose, and the record needs more than one kind of market before it means anything. Glimps has run on a Mac mini since September, and the current simulation began in November. Building a system whose main answer might be “don’t do this” was the point.
Using it found what the tests hadn’t. The brief was describing an account I no longer had, and one long brief came back with its lesson blank. Both were fixed the day they showed, and each now has a check of its own.