Card Game

The Forest of Humbaba is a strategic card game for 2–4 players set in ancient Mesopotamia. It is both a standalone tabletop experience and a research instrument for exploring the Tragedy of the Commons, the Seneca Effect, and the paradoxes of adaptive behaviour. Players take the roles of rival powers vying for the precious cedar of the sacred forest — but the forest is finite, its health fragile, and its true depth uncertain.

---

The Game Concept

The card game translates every mechanical element of the BDPD Platform into the language of a physical tabletop experience. Players draw cards from personal decks, play them simultaneously, and harvest cedars from a shared Forest Deck. The forest regenerates each turn, but its capacity depends on its Health — a track from 0 to 10. If the forest collapses, no one wins. The winner is whoever has the most cedars — but only if the forest survives.

"The card game builds intuition. The simulation explains why." — BDPD Project Mantra

---

Relationship to the Platform

Every mechanical element has a direct analogue:

BDPD Platform	The Forest of Humbaba
Commons stock \(S_t\)	Forest Deck — visible cedars
Hidden reserve	Box Reserve — concealed pool feeding regeneration
Regeneration rate \(r\)	Forest Health (0–10). Each turn, \(\lfloor H/2 \rfloor\) cedars move from Box Reserve to Forest Deck
Stochastic collapse (Forest Die)	Physical Forest Die (d6) — when the Forest Deck is empty and harvest is demanded, roll for each excess unit
Seneca Effect (regen shock asymmetry)	Seneca Clip — triggered when Health drops ≥2 in one turn; next turn's regeneration reduced by 2
Harvest capacity	Capacity tokens — each adds +1 to every harvest
The Mule (strategy override)	Stranger-King dual mode — cooperative phase accumulates Patience; defection converts them into a devastating burst
Gate+Rank victory	Survival prerequisite — collapse → no winner; otherwise most cedars wins
Observation noise	Simulator only — --forest-noise injects noise for LLM agents

The mapping is exact but not redundant. The platform's stochastic collapse and Seneca asymmetry are mapped to distinct mechanisms: the Forest Die handles threshold uncertainty, while the Seneca Clip handles the asymmetric regen-shock dynamic. They are complementary.

---

Three Epistemic Functions

The card game serves three purposes that the computational platform cannot:

1. Behavioural validation. Do human players exhibit the same paradoxes observed in the platform? If so, they are properties of the strategic structure, not artefacts of heuristic design.
2. Generative ground truth. LLM agents reading the same rules as human players enable direct comparison of language-mediated and heuristic decision processes.
3. Tacit knowledge elicitation. The tabletop setting compels players to articulate reasoning about hidden reserves and collapse risk in real time, yielding qualitative data that complements quantitative logs.

Neither instrument alone is sufficient: the platform produces statistical regularities but no intuition; the card game produces intuition but no statistics. Together they form a dual-mode laboratory.

---

Components (Physical Edition)

Component	Quantity	Description
Archetype Decks	1 × 18 + 3 × 20 cards	Warrior-King (18), Temple Keeper, River Merchant, Stranger-King
Event Deck	12 cards	Exogenous perturbations drawn on rounds 3, 5, 7
Forest Deck	30 cards	Visible cedar pool (face down)
Box Reserve	20 cards	Hidden regeneration pool
Forest Health Track	0–10	Starts at 6
Capacity tokens	12	Permanent +1 harvest bonus
Patience tokens	10	Stranger-King resource
Renewal tokens	6	One-time regeneration boost
Forest Die	1 × d6	Stochastic collapse
Defection Token	1	Double-sided: Cooperative / Aggressive

---

Simulation Modes

The Python simulator (cards_ai_play.py) implements all rules for automated play in two modes:

Heuristic Engine (--demo)LLM Mode

All players run the same built-in strategies as the platform — Warrior-King (aggressive), Temple Keeper (conservative), River Merchant (adaptive), Stranger-King (dual-mode). No LLM required. Used for all CT1–CT5 tournament experiments.

One or more players are replaced by language model agents accessed via an OpenAI-compatible API. Supports:

• --api-model / --api-base-url for model selection
• --forest-noise FLOAT for Gaussian noise on the Forest Deck estimate
• --nudge for archetype-specific behavioural nudges
• --lock-defection to prevent voluntary Stranger-King defection

---

The Four Archetypes

Archetype	Strategy	Platform Equivalent	Wins By
Warrior-King 🗡️	Maximises harvest; self-amplifying Capacity	Aggressive	Accumulating Stockpile rapidly — often at cost of collapse
Temple Keeper 🌿	Heals the forest; sacrifices personal gain	Conservative	Surviving to the end with a modest Stockpile while others collapse
River Merchant 📈	Reactive; copies trends; exploits information	Adaptive	Flexibility — but suffers from fatal lag and the Vacuum Effect
Stranger-King 🎭	Dual-mode: cooperative → devastating defection	The Mule	Timing the betrayal for maximum burst without triggering collapse

See Archetypes for deck compositions, key cards, and strategy details.

---

Key Mechanics at a Glance

Mechanic	Effect	Platform Equivalent
Forest Die	d6 roll per excess harvest unit; failure = collapse	Stochastic collapse
Seneca Clip	Health drop ≥2 → −2 regen next turn	Regen shock
Vacuum Effect	Exactly one player harvests 0 → richest gains +1	Reactive < Conservative externality (P8 strategic vacuum)
Capacity tokens	Permanent +1 per harvest	Wealth-scaled capacity
Patience tokens	Stranger-King burst on defection	Strategy override
Events	9 exogenous perturbations on rounds 3, 5, 7	Perturbation engine

See Mechanics for rules, probability tables, and design rationale.

---

CLI flags (cards_ai_play.py)

The card-game simulator exposes the following frequently-used flags. For the full list run python3 agents/cards_ai_play.py --help.

Flag	Default	Effect
--deck1, --deck2	required	Archetype decks (warrior, temple, merchant, stranger)
--cards	required	Path to deck JSON (e.g. cards/cards_v03.json)
--seed	42	Master RNG seed — game is byte-deterministic per (seed, decks, args)
--games	1	Number of matches to play
--max-rounds	8	Cap on rounds per game
--forest-deck	30	Initial visible Forest Deck size
--box-reserve	20	Initial hidden Box Reserve size
--forest-collapse-on	5	Forest Die collapses on this value or higher (1–6). Lower = harsher
--forest-noise	0.0	Gaussian noise σ as a fraction of Forest Deck (used by CT3)
--history-window	3	Past turns shown to the LLM in the prompt (a.k.a. max_history for the SDK)
--demo	off	Verbose per-turn log; equivalent to --verbose plus rich formatting
--nudge	off	Add behavioural nudges to LLM system prompts
--lock-defection	off	Remove explicit DEFECT option from the LLM prompt (Stranger-only)
--temperature	0.6	LLM sampling temperature
--api, --api-model, --api-base-url, --api-key-env	—	Use an external OpenAI-compatible API instead of the local llama-server
--save-dir	logs	Where the per-run subdirectory is created

Reproducibility

Any change to --seed, --forest-collapse-on, --forest-noise, --forest-deck, or --box-reserve produces a different game stream even with the same decks. Keep them pinned when comparing runs.

---

Print & Play

All cards can be generated as print-ready PNGs using the SVG-based pipeline. See Card Generator for usage and Print & Play for assembly instructions.