Pablo Zavala · AI Safety Evaluation · Research Engineering
A tournament poker bot built for the CMU Data Science Club competition, aimed at decision-making under uncertainty against varied and adversarial opponents. The bot reads opponent tendencies from live betting signals, routes among specialist strategies through a bandit-style selector, and spends heavier turn-street search only on high-uncertainty spots. A manifest-driven backtest harness with duplicate-deck reconciliation and frozen release snapshots records every promotion decision against fixed opponents and score weights.
Adaptive poker bot with a documented study workflow: posterior opponent routing, packaged bucket-EV scoring, uncertainty-gated turn search, and manifest-driven duplicate-deck backtests with frozen release snapshots.
Evidence rests on the committed bot and the documented method; the duplicate-deck study scripts that regenerate the comparisons stay outside the public tree, a public headline benchmark against the live tournament field remains pending, and the opponent models stay heuristic and statistical.
Role: Pablo: team bot strategy plus the reproducible experiment workflow, from opponent model to promotion gate.
| Axis | Sample | Evaluator | Result | Limitation |
|---|---|---|---|---|
| Opponent regime coverage | Weighted live-field gate spanning passive, balanced/selective, phase-shifting, classifier-trap, and hyper-aggressive proxies | Per-opponent reward-per-hand blended by fixed manifest weights | A single weighted score ranks each candidate across the whole proxy field, so a lucky slice against one style stays visible rather than hidden. | Proxy opponents approximate the field; a public tournament benchmark stays pending. |
| Variance control | Duplicate-deck reconciliation, seed 20260321 | Matched-deck reward-per-hand delta versus the frozen anchor | Baseline and challenger replay identical decks, so the comparison isolates policy skill from deal luck. | Matched decks shrink variance yet still cover a bounded sample. |
| Promotion discipline | Every candidate before it reaches production | Duplicate-deck gate plus weighted live-field gate | A challenger earns promotion only after it clears both gates, passes the engine smoke test, and holds the time budget. | Short screens can mislead, so the harness prefers longer confirmation runs. |
The bot earns its place through a documented decision stack, meaning signal-weighted opponent reads, bandit-style mode routing, and uncertainty-gated search, rather than a single tuned threshold.
Every promotion runs through a manifest that fixes opponents, hand counts, and score weights, then writes a weighted comparison; the committed strategy and workflow notes document how that gate works, though the study scripts themselves stay outside the public tree.
Read the harness as internal self-play against heuristic proxy opponents; treat the reward-per-hand figures in the lab log as screening signal, and await a public tournament benchmark before reading them as field strength.
Tournament poker punishes a static strategy: opponents range from passive stations to hyper-aggressive maniacs, and some shift style mid-match to bait a fixed counter.
Pablo built a layered bot on the club engine plus a study workflow that stages challengers in a separate lane, so production code stays stable while experiments run.
The opponent model weights recent betting signals by street, pot size, and continue cost, flags instability and cheap-signal risk, and feeds a bandit-style selector that routes among solid, pressure, trap, showdown, robust, counter, equity, and predator modes. A hybrid challenger adds posterior regime routing, packaged bucket-EV scoring across five action buckets, and a higher-compute extension of the uncertainty-gated turn search the production bot already runs. Promotion runs through manifest-driven studies with duplicate-deck reconciliation.
The frozen anchor held as the production default because each challenger, including the richer hybrid stack, failed to beat it on the longer weighted gate; the lab log records that verdict alongside the artifact paths.
Evidence stays internal self-play against heuristic proxies over short-to-medium samples; a public tournament headline number remains pending.
The committed strategy write-up, workflow doc, and running lab log document the routing stack, the promotion gates, and each accept-or-reject decision; the duplicate-deck study scripts that regenerate the comparisons stay outside the public tree.