Starter playstyle sim audit

Date: 2026-07-07
Scope: All nine live starters across healthcare, energy, and AI.
Goal: Diagnose where combat_sim.py / run_simulator.py play and deck-building heuristics diverge from each starter’s intended main-line, then patch the sim to behave like a skilled human main.

Executive summary

The run sim previously used one generic greedy heuristic (_card_power + score_value) for every starter. That model is a good fit for burst/hype decks (Venture Capitalist, parts of Tech Founder) but a poor fit for setup→payoff archetypes (Hacker glitch, Plague Doctor bleed, Grid Operator surge, Nuclear Engineer cash flow).

We added scripts/starter_archetypes.py and wired it into:

Win rates are not the optimization target — smarter play can shift balance measurements. Seed-7 / 300-trial full-clear rates before vs after are included below as a sanity check, not a success criterion.

Per-starter diagnosis

The Hacker (control/defensive) — AI

Design: Glitch stacks → energy deny → strike when the enemy is offline.
Old sim failures:
- API Call scored 0 in combat (draw_then_discard_self ignored).
- Glitch Storm had _card_power = 0 — never bought off shop/rewards.
- glitch_payoff valued at flat mult×4, not enemy stacks.
- Play order: Zero-Day → Beta → Obfuscate (finisher before setup).
- +10 Max HP boon (score 50) beat every starter card on a 50-HP body.

After: Obfuscate → Beta → Zero-Day on turn 1; API Call valued; glitch setup cards rank higher in shops; HP boon threshold ~18 for Hacker at 50 max HP.

Remaining gaps: Opponent glitch still doesn’t reduce block in combat sim (pre-existing). 11-card starter retention clog unchanged.

The Venture Capitalist (hype/aggro) — AI

Design: Stack hype fast, cash out with payoffs, pivot before the bubble bursts.
Old sim: Largely aligned — raw _card_power favours Model Collapse, Singularity Event.
Old sim failures: HP boon still competes with hype builders on 60 HP; hype payoffs not sequenced after hype gain on the same turn.

After: Hype setup prioritised before payoffs; lighter payoff deferral for hype/aggro; lower HP boon appetite.

The Tech Founder (balanced/scaling) — AI

Design: Cycle deck, chaos value, pivot often.
Old sim: Moderately aligned (draw/hype scored).
Old sim failures: API Call scored 0 in combat; cycle cards undervalued vs burst rares in rewards.

After: Draw / draw_then_discard_self bonuses; archetype reward weights for cycle pieces. Clears roughly stable (+0.7pp on seed 7).

The Nurse (balanced/sustain) — Healthcare

Design: Block + heal institution, slow inevitable win.
Old sim failures:
- Risk Assessment (0-cost block+draw) lost every reward comparison to HP boon.
- block_scale finishers (e.g. Malpractice Suit line) penalised −500 in score_value.
- Intent-aware block helps, but sustain line not rewarded in shops/rewards.

After: block_scale valued when block is stacked; sustain archetype boosts heal/block in rewards; HP boon slightly favoured (×1.12).

Watch: More defensive combat sequencing can slow kills — monitor Family/MidTier pacing.

The Hustler (aggro/glasscannon) — Healthcare

Design: Trade HP for power; Burnout Protocol exit.
Old sim failures:
- Overtime Shift, Backroom Deal, Server Overclock scored 0 (energy key missing).
- Sim played 0-cost enablers last, stalling burst turns.

After: energy valued at ×18 in combat; self_dmg bonus for glass aggro; largest seed-7 gain (+6pp → +10pp after tune).

The Plague Doctor (bleed) — Healthcare

Design: Stack bleed, viral attrition.
Old sim failures:
- bleed_payoff flat heuristic; setup vs payoff play order wrong.
- Bleed apply cards undervalued in rewards vs HP boon.

After: Stack-scaled bleed payoffs; bleed setup bonus when stacks low; bleed keys in shop synergy.

The Grid Operator (defensive/sustain) — Energy

Design: Surge stack → explosive combo; capacity building.
Old sim failures:
- surge_consume payoffs not gated on surge stacks.
- Surge gain undervalued in rewards vs HP boon.
- Raw damage cards favoured over surge engine in shops.

After: Surge setup/consume sequencing; defensive/sustain shop triage when low HP.

Watch: Surge setup priority can delay lethal turns — seed-7 clear dipped; may need enemy HP retune separately.

The Nuclear Engineer (scaling/risk) — Energy

Design: Survive early, scale cash flow for infinite energy.
Old sim failures:
- Engine cards (Mainframe Boost, Spark Plug) lose every reward fight to HP boon (50).
- Sim doesn’t model “take weak early cards, survive, scale.”

After: Higher cashflow_gain reward weights; lower HP boon (×0.7); scaling shop triage when low HP.

Watch: Still weak on Family boss in seed-7 — likely needs balance pass on Character_MrMrsWatts more than more sim tweaks.

The Wildcatter (aggro/surge) — Energy

Design: HP trade, combustion, zero-energy turns.
Old sim failures:
- Overclock / self-damage enablers scored 0.
- combustion_payoff flat; combustion setup not sequenced.
- Aggro lines deprioritised vs safe HP boons.

After: self_dmg combat bonus; combustion stack payoffs; aggro attack priority boost.

Universal sim bugs fixed (all starters)

Seed-7 / 300-trial full-clear rates (before → after)

Pooled: ~12.9% → ~11.4% on seed 7. Shifts are expected when the sim stops “HP boon spamming” weak runs into bosses.

Hacker pick-pattern shift (illustrative)

Files changed

• scripts/starter_archetypes.py — new archetype profiles and scoring helpers
• scripts/combat_sim.py — stack-aware score_value, combat_play_score, style_label on Combatant
• scripts/run_simulator.py — archetype rewards, events, removals
• scripts/shop_decision.py — archetype shop triage + synergy keys

Recommended follow-ups

1. ~~Combat sim: Model opponent glitch reducing block / discard quality (helps Hacker fairness vs live game).~~ Done — _gain_block() applies CombatManager.CalculateFinalBlock Glitch penalty to all block sources.
2. Balance: Starters whose sim got smarter but slower (Grid Op, Nurse, NE) may need enemy HP trims — separate from sim work.
3. ~~Regression: Add unit tests for combat_play_score ordering on canonical hands per archetype.~~ Done — scripts/test_archetype_play_order.py (wired into test_combat_regression.py).
4. Hacker starter: 10-card retention fix still worth considering independently of heuristics.

Update (2026-07-07): Glitch block + play-order tests

Glitch block fix

CombatManager.CalculateFinalBlock applies -1 block per Glitch stack to all block gained. The sim now routes every block gain through Combatant._gain_block(), including totem block, card block, and dynamic block effects.

Play-order regression tests

scripts/test_archetype_play_order.py locks canonical sequencing for:

Seed-7 clear rates after glitch block (300 trials)

Glitch block helps any deck that glitches enemies (Hacker, AI control picks). Hacker clear rate is still low because Family/MidTier bosses (Mom & Pop, VP) outscale a 50-HP control body even when the sim sequences correctly.

Remaining sim inaccuracies (priority order)

Decks that still struggle with high-skill sim play

These are the starters where even the improved sim rarely full-clears on seed 7 — pointing to balance/enemy tuning, not just AI gaps:

Doing well with smart sim: Hustler (~32%), Plague Doctor (~19%), Nurse (~18%), Tech Founder (~16%) — sim now plays enablers, bleed setup, sustain, and cycle lines correctly.

Next balance levers (not sim): Hacker 10-card deck + Family boss HP; Watts/Founders/VP trims; NE early sustain card or Watts HP.

Update (2026-07-07): Sim/live-game parity pass

Combat mechanics now mirrored

Run/meta mechanics now mirrored

New files

• scripts/sim_parity.py — upgrade paths, combat deck flattening, hype helper
• scripts/test_sim_parity.py — 10 parity unit tests (wired into test_combat_regression.py)

Seed-7 / 100-trial clear rates (post-parity)

Surge-as-energy and hype multiplier notably help energy/hype starters; Hacker ticked up from better economy modeling without enemy changes.

Remaining sim gaps (lower priority)

Seed-7 / 300-trial clear rates (post parity pass #2)

Upgrade rolls and cursed-totem modeling help energy/surge starters most (Grid Op, Wildcatter). Tech Founder/VC drops likely reflect curse-totem downside + capital drain when Cursed Doll is owned, and fewer “free” HP-padding paths as upgrades replace raw picks.

Balance outliers (skill-aligned sim, seed 7)