nsba-markets • 2026-08-04

Can a bot destroy you at Buzzer? — the AI ceiling investigation

TL;DR

You asked me to run the gauntlet overnight and try to make the bot play meaningfully better —

ideally reproduce your wins in sim and beat them. Here's the honest result: **the bot already

crushes every sim opponent I can build (an all-in category-stacker loses 95%), so the sim cannot

reproduce you beating DeepBlue.** I measured all three decision axes — play, category-pick, and

draft — and only ONE has real head-room: the category pick (DeepBlue+, already shipped, beats

DeepBlue 64.8%). Play and draft are both at the sim ceiling (draft head-room is literally negative,

n=1500). **You were right that the game isn't RNG-capped — but the reason you beat the bot isn't a

fixable in-sim deficiency. It's that you're a stronger, more adaptive opponent than anything the sim

can generate.** To actually build a bot that beats *you*, I need per-move logs of you playing

DeepBlue+. Go play it (it's live in the picker) — that's the one experiment I can't run without you.

What I was asked to do

> "Run the gauntlet and then simulate games against me or something else. Give it the pgn and ask

> what it would do — see if it plays meaningfully different or drafts different teams. If it's still

> bad try to improve it again and upload everything onto share and message me on discord."

Context: you'd just beaten DeepBlue three times in a row, lopsidedly (176-130 biology 6-2; 172-86

math 9-1; +1 more), by stacking and steering one category. Earlier in the week I'd concluded the

game was "variance-capped, ~1-3pp head-room" — and your 100% win rate made that claim look absurd.

So the real question this session: is that conclusion wrong, and if so where's the head-room?

What I found (the three axes)

The earlier "1-3pp head-room" was measured wrong: it only ever optimized posture (boost/pull/

decline) while holding the category pick and the draft greedy. That's measuring inside a

strategy space that *excludes the thing you actually do to win*. So I re-measured the two unsearched

axes directly.

1. Category pick — the ONE positive lever (shipped as DeepBlue+).

The winner names the next category. Greedy bots just name their strongest held category; DeepBlue+

*searches* it — rolls the engine forward for each category it still holds a pick-card for and picks

the one with the best win-probability. Results:

• Beats DeepBlue 64.8% (confidence interval excludes 50 — real, not noise)
• Beats Magnus 59.5%, MegaSanj 54%
• Now the apex bot on the ladder, seeded Elo 2620

This is the first genuinely positive AI result of the whole investigation. It's live — you can pick

it in the bot dropdown.

2. Draft — ≈0 head-room, actually NEGATIVE (n=1500, well-powered).

A subagent isolated the draft (both seats run the identical in-game engine; only the deck differs, so

a head-to-head win rate *is* the draft's value) and tried five strategies against the shipped

win-rate-prior draft:

Nothing beats the shipped draft. The equal-exposure win-rate priors already encode deck value so

well that even a search explicitly optimizing decks to beat them couldn't. Deckbuilding is near-

ceiling too. (So no, the bot doesn't need to "draft different teams" — its draft is already ~optimal.)

3. Play / posture — near-ceiling (measured earlier).

A perfect *peeking* best-responder (an upper bound on exploitation — it literally rolls your exact

policy forward as its opponent model) beats the strong bots by only ~1-3pp. Opponent-posture modeling

came out ~-5pp: the board is public so the search already exploits a weak field, and AGGRESSIVE is a

convert *bonus*, so "over-pressing" isn't even a leak to exploit. A deep-learned value net was a

better *predictor* (log-loss 0.34 vs 0.40) but a dead heat in *play* (50.5%) — the simple

margin / σ√(questions_left) leaf already carries all the rankable signal.

Why the sim can't reproduce your win (the actual wall)

I built your winning strategy as a sim bot: an all-in single-category stacker that fields and steers

that category every pick. Then I ran DeepBlue and DeepBlue+ against it, 80 games per category:

                 win rate   margin   farm (rds its category resolved)
DeepBlue  vs mono   94.9%    +84.8    5.9
DeepBlue+ vs mono   95.2%    +78.8    5.6

The naive stack loses 95%. It successfully farms its category ~6 rounds a game — and still gets

blown out, because an all-in deck gasses out on the *other 18* rounds. So stacking alone is not why

you win. You keep a *competitive, balanced* deck AND adaptively steer a category the bot is weak in,

reading the specific matchup. No sim opponent — stacker, HumanSim, specialist, best-responder — models

a strong adaptive human. **The bot beats every proxy I can build, so I can't tune it to beat you in

sim: there's no sim opponent standing in for you that it doesn't already crush.**

One structural note that matters: you cannot "deny a farm" with pick-search. The farmer farms by

winning *their own* pick rounds — the bot only picks a category when *it* wins. The only anti-farm

lever is winning the contested category outright (field/contest/draft), and the search already presses

exactly where win-probability says to.

What succeeded

• DeepBlue+ pick-search — merged + deployed (8f97364). The apex bot; beats DeepBlue 64.8%.
• DeepBlue added to the gauntlet — ranks #2 at 82.7%, dead even with Magnus/MegaSanj (f947785).
• DeepBlue+ available in the live bot picker — go play it.
• Draft head-room definitively closed — clean negative, n=1500 (121e572, on a subagent branch).
• Mono-stacker crux sim — proves a naive stack loses 95% (079e5fd).

What failed or is a dead end

• MCTS / CFR / MCCFR — rejected (a no-context red-team independently called MCCFR a no-go:

abstraction pathology on continuous stamina/score, Nash≠exploit, variance double-bind).

• Deep-learned value leaf — better predictor, dead heat in play. Wrong layer.
• Belief-steering + field-search (DeepBlue+ variants) — washed out / hurt; shipped OFF.
• The whole "improve the bot until it beats the user in sim" loop — not runnable: the bot already

beats every sim opponent, so there's no in-sim gradient toward beating you.

Current state

Everything's merged to main and deployed except the draft harness (a clean negative — left on its

subagent branch worktree-agent-ac2ec3a4aeb837396, not worth polluting the ladder with a losing

bot). DeepBlue+ is the live apex bot. Backend suite green (1057 passed).

Recommendations (priority order)

• Play DeepBlue+ a bunch and let me log every move. This is the one experiment I can't run

without you. It's the only path to a bot that beats *you specifically*: per-move data → an online

opponent model (find what you do that it can punish) or an offline best-response to your

tendencies. Right now I'm modeling a phantom.

• Accept the ceiling for the *generic* bot. On every axis I can measure in sim, it's at or near

optimal. Fancier engines (search depth, CFR, value nets) are all measured flat — don't fund them.

• If you want the bot to *feel* harder without being smarter: give it a small information edge

(e.g. it sees a noisier version of your bench) — that's a design lever, not an AI one, and it's

the honest way to make a bot that pressures a top human.

Commit log

079e5fd test(card game): mono-stacker crux sim — both bots crush an all-in stacker ~95%
8f97364 merge: DeepBlue+ — searches the category pick, beats DeepBlue 64.8%
0a36e9c tune(card game): DeepBlue+ — pick-search is the measured win; Elo 2620
f947785 feat(card game): add DeepBlue to the gauntlet (ranks #2, 82.7%)
0091dae feat(card game): DeepBlue+ — search the category pick + fielding
456387c merge: DeepBlue win-probability leaf + (off-by-default) opponent model
121e572 (subagent branch) draft head-room harness — clean negative, n=1500