# Vexa 1.1: results after the Strategy Lab audit

Evaluated September 6, 2026 (UTC). Vexa is an authored tactical opponent, available in Play vs Bot and Watch Bots. The lab supports developing, comparing, and submitting new strategies; it does not make the game learn automatically from every match.

## What changed

Vexa considers immediate combat trades, preserves gems when sacrificing them has no useful payoff, and avoids pointless self-targeted Royal Sacrifice. The audit added a bounded check for obvious lethal counterattacks: it now holds its last defender in the reproduced losing endgame, while still taking attacks that win immediately. This check uses the public board and does not predict future card plays.

## Fresh independent evaluation

The packaged `first-challenger@1.2.0` strategy wraps the same updated TacticalAgent as Vexa 1.1. We evaluated its exact portable JavaScript bundle in an isolated QuickJS runtime. Each opponent received 512 fresh deal groups, played across both seats and both first-player choices. Policies received player-visible information and independent randomness.

| Opponent | Wins | Losses | Draws | Score |
| --- | ---: | ---: | ---: | ---: |
| Existing Greedy (Boldrick) | 2,048 | 0 | 0 | 100% |
| Random (Pip) | 2,048 | 0 | 0 | 100% |
| Updated TacticalAgent mirror | 1,023 | 1,023 | 2 | 50% |

All 6,144 games completed with zero policy execution failures. Draws count as half a win. The conservative 95% score interval against Greedy was 93.998%–100%, calculated over independent deal groups. The evaluation limited decisions to 100 ms and sandbox memory to 32 MiB. Browser hardware varies.

## Why the old baseline was too easy

A separate audit of the original release repeated 128 games against Greedy, then repaired only its resource scoring for gem sacrifice, Royal Sacrifice, and Fireball. Original Vexa won 128/128 against unmodified Greedy but 87/128 (67.97%) against the repair on the same deals and seat/starting-player schedule. The repair also scored the full reduced menu instead of Greedy's 600-action sample; this was a combined diagnostic, not a controlled proof about one individual rule.

The original perfect score mostly exposed poor resource decisions in the old opponent. In a retained replay, Greedy ended with no gems and five uncast cards after consuming 13 gems; Vexa retained six. This was ordinary gameplay, not a forfeit or scoring error.

That audit evidence belongs to the original source snapshot and has not been relabeled as a test of Vexa 1.1. It explains why wins against Greedy and Random alone are insufficient. The general-opponent import gate now also requires an observed score of at least 50% against the current Vexa snapshot. That is a minimum gate, not statistical proof of superiority.

## Remaining work

These results establish improvement over the previous baseline. They do not establish strength against experienced humans or diverse future community bots. The mirror matchup is a consistency check, not a separate strategic challenger. Vexa still lacks long-horizon planning and search over hypothetical hidden information. Its counterattack check catches obvious public threats, with optimistic allowances for blocks and kills, rather than exhaustively solving future combat.

Useful next experiments include preserving straights, timing Aces and removal, planning Jack merges, and testing when resource conservation becomes too passive. Compare each idea against Vexa and saved alternatives; retain counterexamples and ablations in the report. Contributions remain subject to manual review.

## Reproducible identity

- Strategy: `first-challenger@1.2.0`
- Browser opponent: `superjack-vexa@1.1.0`
- JavaScript SHA256: `b76de931892035aa1503f88d80d519b1853c67c8bf3b7282bdb830c0757dbe49`
- Engine/agents SHA256: `5fff9e35cc91650d4a8bf28f4116742cfa5472a2f6d51eeeea598f90294deed9`
- Rules SHA256: `ce6c40528bf2f34d965dadd84721a3a4ebc0ac13fd54e1f25c70bd0c13e37a66`

Strategy Lab 0.2.1 contains this exact simulator snapshot, including TacticalAgent and its regression tests. The complete evaluation, replays, and original ablation are retained in the first-challenger campaign. See the [research guide](https://superjackthegame.com/research.md) to develop and submit a competing strategy.