review of final term of game #77
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
ngl the game was kinda easy... idk maybe not much you can gleam from a single turn frame but maybe something will be insightful
Reading the frame, three things jump out before I touch any code:
1. The loan market has a dominant lowball strategy, and it's yours. You bid 2.0000% and won at 3.6733% on all three loans. That gap is Vickrey working as designed (winner receives the SECOND-lowest rate, which is above your own bid — the runner-up's true valuation). Consequence: your bid only decides whether you WIN, never what you earn, so the optimal bid is always the floor. Every AI bids its honest valuation (2.87–3.67%), you bid 2%, you sweep at their rates. If you'd bid 0.1% on all 10 loans you'd have won all 10 at 2.87–3.67%. That one asymmetry is doing a lot of the 'easy'. (It also means the AI floor-bidders can never undercut you profitably — undercutting only changes the second price, not your payout.) This may need a ruling from you: keep Vickrey, or make the winner stand by its own bid (pay-your-bid), or cap the winner's payout at its own bid. I'll put numbers to each option.
2. The deposit market was abandoned by every AI. All seven AI banks posted 0.0000% while you paid 0.6% and captured ~100k of the 109.6k pool per turn. Paying 0.6% (~2.2k/turn) to monopolize 100k/turn of funding is the best trade in the game, and every AI left it on the table. Your 355k deposit base vs their 93k is the actual 'easy'. I need to find out WHY they're all at zero — which brains these were, and whether their rate logic degenerates at this config.
3. The opposition died before you met it. Five AIs bankrupt by turn 15; the two survivors are passive (tiny loan books, zero rate bids). The aggressive funding brains appear to have self-destructed early, leaving the parking brains — that's why the endgame felt empty. The acquisition you made for 0 on Bank 2's −3,147 book was correct (second price 0, the other bidders rationally bid 0 for a negative net).
So my plan: reconstruct this config (8 players, ~d30 die, new 16k, churn 15%, op 2,000, loan terms to 27), run it headless with AI-only rosters across seeds, and measure exactly who dies when and what rates everyone posts. Then I'll come back with the numbers and fix what the data says is broken — brain logic where it's a brain bug, engine ruling where it's the auction.
Ruling received - engine stays as-is, the AI must get good enough to notice the cheese. Agreed, and the data backs the "full overhaul" conclusion. What the reproduction measured, so v0.5.0 starts from numbers and not vibes:
Your exact config, 20 seeds, AI-only (current master brains):
So at your play config the human's real opposition is roughly two-thirds dead by turn 10, one-third parked, and 40% of the time the AI market fully collapses on its own. Eating the leftovers for $0 at auction is the rational endgame of that - the auction isn't the problem, the corpses are.
The two structural causes, named:
What I'd put in the v0.5.0 overhaul brief (tracked in #26): distinct per-brain deposit models with at least one structural leader; cheese detection (second-price awareness, respond to being ratcheted); term-compounded default pricing (the flat die-rank underprices the 27-turn loans at this config - that's what's killing the three lenders); and a survival floor so the table stops wiping out 40% of the time.