diff --git a/environment.md b/environment.md
new file mode 100644
index 0000000..cba57d3
--- /dev/null
+++ b/environment.md
@@ -0,0 +1,216 @@
+# Game Design Document: **Honey Heist: Battle of the Bears**
+
+---
+
+## 1. Concept Paragraph
+
+**Honey Heist: Battle of the Bears** is a deterministic, turn-based strategy game where two rival bears compete in a forest clearing to collect the most honey from a shared hive before it is depleted. Each bear must decide whether to **forage**, **defend**, or **steal** on its turn, balancing risk against reward. The environment features a finite supply of honey, and both bears act in partial opposition — every move affects the shared resource and their honey stores.  
+This design is **entirely unrelated to any negotiation or trading game**. It focuses instead on resource competition and strategic timing in an original woodland theme.
+
+---
+
+## 2. Roles and Win Condition
+
+- **Players:** Two rival bears: **Bear A** and **Bear B**.
+- **Objective:** End the game with **more honey than the opponent** when the hive is empty or the maximum number of turns is reached.
+- **Win Condition:**
+  - **Win:** A bear has more honey than the other at the end of the game.
+  - **Loss:** A bear has less honey than the other when the game ends.
+  - **Draw:** Both bears have equal honey at game end.
+
+---
+
+## 3. Turn Structure and Determinism
+
+- The game proceeds in **alternating turns** beginning with Bear A.
+- On each turn, the current player submits **one valid action token**.
+- Each turn’s outcome is **deterministic**, derived exclusively from the game state and player actions.
+- The game lasts until either:
+  1. The **hive’s honey** reaches **0**, or  
+  2. A **turn limit** (default = 10 rounds per bear, 20 turns total) is reached.
+- A fixed **seed** ensures deterministic replay: the hive’s starting honey and any numeric parameters (like attack/defense outcomes) are seeded and reproducible.
+
+---
+
+## 4. Action Grammar (Machine-Parseable)
+
+### Allowed Player Actions
+
+| Action Token | Pattern Format | Description |
+|---------------|----------------|--------------|
+| `[Forage:X]` | `^\[Forage:(?:1|2|3)\]$` | Gather **X** units of honey from the hive (reduces hive honey by X). |
+| `[Defend]` | `^\[Defend\]$` | Protect your stored honey; blocks part of a steal attempt this turn. |
+| `[Steal:X]` | `^\[Steal:(?:1|2|3)\]$` | Attempt to steal **X** honey from the opponent’s store (only effective if they did not defend). |
+
+### Examples
+
+**Valid:**
+- `[Forage:2]` → Forages 2 honey from hive.
+- `[Steal:3]` → Attempts to steal 3 honey from the opponent.
+- `[Defend]` → Adopts a defensive stance.
+
+**Invalid:**
+- `[Forage:5]` → Invalid, X must be within 1–3.  
+- `[Steal honey]` → Invalid format; must use integer param.  
+- `[Hide]` → Invalid token; not part of the grammar.
+
+---
+
+## 5. Game State Schema
+
+```json
+{
+  "turn_number": 7,
+  "current_player": "BearB",
+  "hive_honey": 12,
+  "max_turns": 20,
+  "players": {
+    "BearA": {
+      "stored_honey": 9,
+      "last_action": "[Forage:2]",
+      "defending": false,
+      "score": 9
+    },
+    "BearB": {
+      "stored_honey": 10,
+      "last_action": "[Steal:3]",
+      "defending": false,
+      "score": 10
+    }
+  },
+  "history": [
+    {"turn":1,"actor":"BearA","action":"[Forage:3]"},
+    {"turn":2,"actor":"BearB","action":"[Defend]"},
+    {"turn":3,"actor":"BearA","action":"[Steal:2]"}
+  ],
+  "winner": null,
+  "draw": false,
+  "seed": 12345
+}
+```
+
+---
+
+## 6. Initialization Rules
+
+- **Hive honey** starts between **15 and 20**, seeded deterministically.
+- **Stored honey** for each player starts at **0**.
+- **Turn number** begins at **1**, with **BearA** going first.
+- A random seed determines only initial hive honey; all later effects are deterministic.
+- On **reset()**, both players receive initial observations describing:
+  - Starting hive size
+  - Their honey score (=0)
+  - Available action grammar
+
+---
+
+## 7. Validation and Error Handling
+
+### Illegal Actions
+The environment validates the extracted content (from within `\boxed{}`).
+
+**Invalid Move Reasons:**
+1. Action string **empty or malformed** → `"Invalid format, must use [Forage:X], [Steal:X], or [Defend]."`
+2. **X out of range** for `[Forage:X]` or `[Steal:X]` → `"Illegal quantity, X must be 1–3."`
+3. Attempt to forage more honey than exists in hive → `"Not enough honey in hive."`
+4. Attempt to steal more honey than opponent has → `"Opponent has insufficient honey."`
+5. Action submitted after terminal state → `"Game is already over."`
+
+When validation fails, the environment triggers `set_invalid_move(reason)` and ends the turn without effect.
+
+---
+
+## 8. Terminal Conditions and Scoring
+
+### Checked after each pair of turns (full round):
+
+1. **Hive depletion:** if `hive_honey <= 0`, game ends.
+2. **Turn limit:** if `turn_number > max_turns`, game ends.
+3. **All honey gathered:** if total honey = 0 (hive + players), end early.
+
+**Scoring:**
+- Each bear’s score = `stored_honey`.
+- **Winner = bear with higher stored_honey.**
+- **Draw** if equal.
+
+**Tie-breaker:** If equal and turns remain, continue until max_turns; if still equal at end → draw.
+
+---
+
+## 9. Player Prompt Specification
+
+Each player receives a contextual prompt describing their role and the rules:
+
+---
+
+### Prompt Outline
+
+```
+You are a hungry bear competing for the last honey in the forest.
+- Your goal: End the game with more honey than your rival.
+- Each turn, choose ONE of the following actions:
+  [Forage:X]  Gather X units (1–3) from the hive.
+  [Defend]    Protect your honey from theft this turn.
+  [Steal:X]   Steal X units (1–3) from your rival if they do not defend.
+
+Game facts:
+- Hive honey remaining: {hive_honey}
+- Your stored honey: {player_honey}
+- Rival stored honey: {rival_honey}
+- Turn {turn_number} / {max_turns}
+
+Format rule:
+State your reasoning briefly, then put your final action in the following format at the end:
+
+"Put your final answer within \\boxed{{}} at the end of your response."
+
+Example valid response:
+I will risk a quick grab of honey before the hive gets low.
+\boxed{{[Forage:3]}}
+
+Example invalid response:
+\boxed{{Forage 3}}  <-- Must include brackets and colon.
+```
+
+---
+
+A helper `_extract_answer_content(self, action: str) -> str` will strip the outer `\boxed{}` to obtain the core action token for parsing and validation.
+
+---
+
+## 10. API Mapping Plan
+
+### `reset(seed)`
+- Initialize `game_state` fields as specified.
+- Apply seed to determine initial `hive_honey`.
+- Set both `stored_honey` = 0, `defending`=False.
+- Return the first observation for **BearA**.
+
+### `step(action)`
+- Extract content from player’s input using `_extract_answer_content`.
+- Validate; if illegal, set invalid move and skip reward changes.
+- Apply deterministic resolution:
+  - `[Forage:X]`: Reduce hive honey by X, add to player’s store.
+  - `[Defend]`: Set defending=True.
+  - `[Steal:X]`: If opponent.defending=False, transfer min(X, opponent.honey) to current player.
+- Reset defending flags at the end of round.
+- Increment `turn_number`, switch `current_player`.
+- Check terminal conditions and set `winner`/`draw` if applicable.
+- Produce updated observation for next player.
+
+### `_generate_player_prompt(player_id)`
+- Inserts current state data into the prompt template (see section 9).
+- Explains available actions, format rules, and examples.
+- Used each turn to query model actions in deterministic play.
+
+---
+
+## 11. Copy-Check Against Example
+
+This design—
+- Uses **bears, honey, and forest resources**, not trade, offers, or negotiation.
+- Has **unique resource names**: `hive_honey`, `stored_honey`, `defending`, etc.
+- Defines **original objectives** (collect more honey) and **distinct action grammar**.
+- Uses distinct theme and prompt text centered on **competitive foraging bears**.
+  
+All terminology and `game_state` keys are purely original to *Honey Heist: Battle of the Bears*.
\ No newline at end of file