Decision traces
Every bet can explain itself, from the moment of the decision, not after the fact.
What a trace holds
Every decision, executed or refused, records four things at the moment it is made.
- What it saw. The game, the payout, the computed house edge, the bankroll, the day’s PnL, the streak coming in.
- What the policy checked. Every dial by name, with the live numbers on both sides: "Per-bet cap: 0.32 ≤ 0.50 AVAX", pass or fail, one line per rule.
- Why this size. The sizing logic in the agent’s own words, tied to its temperament and the numbers above.
- The transaction. Hash, timestamps, settlement, and the final net, colored honestly.
Recorded at decision time, never reconstructed
The trace is written when the decision happens, before the result exists. When the agent explains itself later, it narrates that stored record and nothing else.
The explanation call is deliberately starved: the model gets the stored snapshot, the verdict, and the outcome. No market access, no fresh portfolio, no tools. It cannot discover a better-sounding reason than the one it actually had, because it cannot see anything except what it saw at the time.
Where to find a trace
- In chat. Ask "why" and the last resolved bet comes back with its reason and a card linking the full trace.
- In the feed. Every bet and result row opens into its trace. LIMIT rows carry the named rule that fired.
- On positions. Each position shows the trace that opened it, so the entry logic is one tap from the entry.
Reading a trace, step by step
Start at the top: what it observed. This is the world as the agent saw it, including the house edge stated plainly. If the observation looks thin, the decision was thin.
Then the checks. Four rules, each with the live numbers at that moment. This block is the policy engine showing its work: you can recompute every line yourself.
Then the sizing reason, which is where temperament shows. Two agents with the same policy and the same table will size differently, and this line tells you why. Last, the transaction itself, so the story ends onchain where anyone can verify it.
Refusals have traces too
A blocked bet produces the same record with a failed check in the middle. Ask about it and you get something like: I wanted 0.8 on Limbo, that would have dipped into the 3 AVAX reserve you set, so it did not happen.
These are often the best reads in the product. They show the exact moment your rules did their job, with the numbers that proved it.
Why this matters
An agent that bets real money and cannot show its work is a slot machine with a face. The trace is what separates an explained decision from a story invented after the result.
It also keeps the agent honest with itself. Its memory of a night is colored by temperament. Its trace of a night is not. When the two disagree, you have the receipts.