diff --git a/berghain/__init__.py b/berghain/__init__.py
new file mode 100644
index 0000000..95492ec
--- /dev/null
+++ b/berghain/__init__.py
@@ -0,0 +1,2 @@
+__app_name__ = "berghain"
+__version__ = "0.0.1"
diff --git a/berghain/__main__.py b/berghain/__main__.py
new file mode 100644
index 0000000..5fe53a5
--- /dev/null
+++ b/berghain/__main__.py
@@ -0,0 +1,3 @@
+from .cli import main
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/berghain/api.py b/berghain/api.py
new file mode 100644
index 0000000..e5fb357
--- /dev/null
+++ b/berghain/api.py
@@ -0,0 +1,50 @@
+from typing import Any
+import requests
+from . import __version__
+from .config import BASE_URL
+
+class APIRequestError(Exception):
+    """Raised when the API request is malformed"""
+
+class APIResponseError(Exception):
+    """Raised when the API returns an error"""
+
+class APIClient:
+    def __init__(self) -> None:
+        self.base_url = BASE_URL.rstrip("/")
+        self.session = requests.Session()
+        self.session.headers.update({"User-Agent": f"kaylee's berghain-cli/{__version__}"})
+
+    def new_game(self, player_id: str | None, scenario: int) -> dict[str, Any]:
+        """
+        Start a new game for this player
+        """
+        if player_id is None:
+            raise APIRequestError("A player ID is required")
+        return self.__do_request("new-game",{"scenario": scenario, "playerId": player_id}, "gameId")
+
+    def decide_and_next(self, game_id: str, person_index: int, accept: bool | None = None) -> dict[str, Any]:
+        """Submit a decision for the current person and get the next one"""
+        params: dict[str, Any] = {"gameId": game_id, "personIndex": person_index}
+
+        if person_index > 0 and accept is None:
+            raise APIRequestError("You must specify whether to accept or reject this person")
+
+        if accept is not None:
+            params["accept"] = str(accept).lower()
+
+        return self.__do_request("decide-and-next", params, "status")
+
+
+    def __do_request(self, endpoint: str, params: dict[str, Any], check_for: str | None = None) -> dict[str, Any]:
+        resp = self.session.get(f"{self.base_url}/{endpoint}", params=params, timeout=10)
+        if not resp.ok:
+            raise APIResponseError(f"HTTP {resp.status_code}: {resp.text}")
+        try:
+            data = resp.json()
+        except ValueError as e:
+            raise APIResponseError(f"Invalid JSON in response: {resp.text}")
+        if check_for is not None and check_for not in data:
+            raise APIResponseError(f"Malformed response (key \"{check_for}\" not found): {data}")
+
+        return data
diff --git a/berghain/cli.py b/berghain/cli.py
new file mode 100644
index 0000000..d4bac7e
--- /dev/null
+++ b/berghain/cli.py
@@ -0,0 +1,386 @@
+from __future__ import annotations
+
+import os
+import sys
+import signal
+from collections import deque
+import json
+from typing import Optional, Deque
+
+from docopt import docopt
+
+from .core import BerghainClient
+from .policy import load_policy, PolicyLoadError, default_policy_path
+from .config import VENUE_CAP, REJECT_CAP
+from .state import GameStatus
+from .ui import ProgressiveRenderer
+
+USAGE = """berghain
+
+Usage:
+  berghain set-player <player>
+  berghain new [--scenario=<n>]
+  berghain play [--interactive] [--verbose]
+  berghain step [--verbose]
+  berghain list
+  berghain status [<idprefix>]
+  berghain switch <idprefix>
+  berghain set-policy (default | <file>)
+  berghain check-policy <file>
+  berghain (-h | --help)
+  berghain --version
+
+Options:
+  -h --help       Show this screen.
+  --version       Show version.
+  --scenario=<n>  Scenario number (1, 2, or 3). [default: 1]
+  --interactive   Pause for Enter between steps.
+  --verbose       Send raw transcript lines to stderr (policy debug prints too).
+"""
+
+TEMPLATE_POLICY = """# policy.py (user template)
+# Write your own decision logic. Must define:
+#   decide(attributes, tallies, mins, admitted_count, venue_cap, **kwargs) -> bool
+#
+# Hints:
+# - 'attributes' are booleans for the current person.
+# - 'tallies' counts already-admitted attributes.
+# - 'mins' are required minimum counts per attribute.
+# - You may optionally accept: p (relative frequencies), corr (correlation matrix).
+
+def decide(attributes: dict[str, bool],
+           tallies: dict[str, int],
+           mins: dict[str, int],
+           admitted_count: int,
+           venue_cap: int,
+           **kwargs) -> bool:
+    # Example starter: admit only if they help an unmet attribute.
+    need = {a: max(0, mins[a] - tallies.get(a, 0)) for a in mins}
+    if sum(need.values()) == 0:
+        return True
+    return any(attributes.get(a, False) and need[a] > 0 for a in mins)
+"""
+
+
+# -------- transcript helpers --------
+
+def _raw_line(entry: dict, mins: dict[str, int]) -> str:
+    idx = entry.get("personIndex")
+    accept = entry.get("accept")
+    attrs = entry.get("attributes", {})
+    need_before = entry.get("remaining_need_before", {})
+    xs = sorted([k for k, v in attrs.items() if v])
+    helped = sorted([a for a in xs if need_before.get(a, 0) > 0])
+    return f"#{idx:>5}  {'ACCEPT' if accept else 'reject':<6}  attrs: {', '.join(xs) if xs else '∅':<24}  helped: {', '.join(helped) if helped else 'none'}"
+
+def _stderr(s: str) -> None:
+    print(s, file=sys.stderr, flush=True)
+
+# -------- signal handling (finish current person, then exit) --------
+
+class _SoftSigint:
+    """Install a SIGINT handler that sets a flag and does not raise KeyboardInterrupt.
+    Use .armed and .disarm to toggle around blocking input.
+    """
+    def __init__(self) -> None:
+        self._old = None
+        self.triggered = False
+
+    def _handler(self, signum, frame):
+        self.triggered = True
+
+    def arm(self) -> None:
+        self.triggered = False
+        self._old = signal.getsignal(signal.SIGINT)
+        signal.signal(signal.SIGINT, self._handler)
+        # Restart interrupted system calls rather than raising:
+        try:
+            signal.siginterrupt(signal.SIGINT, False)
+        except AttributeError:
+            pass  # not available on some platforms
+
+    def disarm(self) -> None:
+        if self._old is not None:
+            signal.signal(signal.SIGINT, self._old)
+            self._old = None
+
+# -------- commands --------
+
+def _print_game_summary(rec) -> None:
+    mins = rec.constraints.mins
+    print(f"Game: {rec.game_id}")
+    print(f"  Scenario: {rec.scenario}")
+    print(f"  Status:   {rec.status}")
+    print(f"  Admitted: {rec.admitted_count}")
+    print(f"  Rejected: {rec.rejected_count}")
+    print(f"  Policy:   {rec.policy_path or '(none)'}")
+    if mins:
+        print("  Constraints:")
+        for a, m in mins.items():
+            have = rec.tallies.get(a, 0)
+            print(f"    - {a}: {have}/{m}")
+    print()
+
+def _ensure_template(path: str) -> None:
+    dirpath = os.path.dirname(os.path.abspath(path)) or "."
+    os.makedirs(dirpath, exist_ok=True)
+    with open(path, "w", encoding="utf-8") as f:
+        f.write(TEMPLATE_POLICY)
+
+def cmd_set_player(player: str) -> int:
+    bc = BerghainClient()
+    bc.set_player(player)
+    print(f"Player set: {player}")
+    return 0
+
+def cmd_new_game(scenario: int) -> int:
+    bc = BerghainClient()
+    if scenario not in (1, 2, 3):
+        print("--scenario must be 1, 2, or 3.", file=sys.stderr)
+        return 2
+    rec = bc.new_game(scenario=scenario)
+    _print_game_summary(rec)
+    print("Tip: attach a policy with `berghain set-policy ./my_policy.py` (required before `play`).")
+    return 0
+
+def cmd_status(idprefix: str | None) -> int:
+    bc = BerghainClient()
+    rec = None
+    if idprefix:
+        rec = bc.gs.get_game(idprefix)
+        if not rec:
+            print("No unique match for that id prefix.", file=sys.stderr)
+            return 1
+    else:
+        rec = bc.gs.current_game()
+        if not rec:
+            print("No current game. Start one with `berghain new`.", file=sys.stderr)
+            return 1
+
+    _print_game_summary(rec)
+    print("Initial conditions:")
+    print(json.dumps(bc.gs.current_game().initial_raw, indent=2))
+    return 0
+
+
+def cmd_set_policy(path: str) -> int:
+    bc = BerghainClient()
+    rec = bc.gs.current_game()
+    if not rec:
+        print("No current game. Start one with `berghain new`.", file=sys.stderr)
+        return 1
+
+    if path.strip().lower() == "default":
+        try:
+            load_policy(default_policy_path())
+        except PolicyLoadError as e:
+            print(f"Default policy error: {e}", file=sys.stderr)
+            return 2
+        rec = bc.set_policy_for_current_game(default_policy_path())
+        print(f"Policy set for game {rec.game_id}: {rec.policy_path}")
+        return 0
+
+    if not os.path.exists(path):
+        print(f"Policy file not found, creating template: {path}")
+        try:
+            _ensure_template(path)
+        except OSError as e:
+            print(f"Could not create template policy: {e}", file=sys.stderr)
+            return 2
+    try:
+        load_policy(path)
+    except PolicyLoadError as e:
+        print(f"Policy error: {e}", file=sys.stderr)
+        return 2
+    rec = bc.set_policy_for_current_game(path)
+    print(f"Policy set for game {rec.game_id}: {rec.policy_path}")
+    return 0
+
+def cmd_check_policy(path: str) -> int:
+    if path.strip().lower() == "default":
+        try:
+            load_policy(default_policy_path())
+        except PolicyLoadError as e:
+            print(f"Invalid default policy: {e}", file=sys.stderr)
+            return 2
+        print("Default policy looks good.")
+        return 0
+    elif not os.path.exists(path):
+        print(f"Policy file not found, creating template: {path}")
+        try:
+            _ensure_template(path)
+        except OSError as e:
+            print(f"Could not create template policy: {e}", file=sys.stderr)
+            return 2
+    try:
+        load_policy(path)
+    except PolicyLoadError as e:
+        print(f"Invalid policy: {e}", file=sys.stderr)
+        return 2
+    print("Policy looks good.")
+    return 0
+
+def cmd_step(verbose: bool) -> int:
+    bc = BerghainClient()
+    rec = bc.gs.current_game()
+    if not rec:
+        print("No current game. Start one with `berghain new`.", file=sys.stderr)
+        return 1
+    if not rec.policy_path:
+        print("This game has no policy set. Use `berghain set-policy ./policy.py` first.", file=sys.stderr)
+        return 2
+
+    last5: Deque[str] = deque(maxlen=5)
+    renderer = ProgressiveRenderer(rec.constraints.mins, VENUE_CAP, REJECT_CAP)
+    renderer.begin()  # reserve and print the block FIRST
+
+    # Soft SIGINT so the current network call finishes
+    guard = _SoftSigint()
+    guard.arm()
+    try:
+        entry = bc.step(rec)
+    except PolicyLoadError as e:
+        print(f"Policy error: {e}", file=sys.stderr)
+        guard.disarm()
+        return 2
+    finally:
+        guard.disarm()
+
+    # Update UI
+    line = _raw_line(entry, rec.constraints.mins)
+    if verbose:
+        _stderr(line)
+    last5.append(line)
+    renderer.render(last5, rec.tallies,
+                    entry.get("admitted_count", rec.admitted_count),
+                    entry.get("rejected_count", rec.rejected_count))
+
+    # If SIGINT occurred during the step, exit now (current person finished)
+    if guard.triggered:
+        _stderr("Interrupted. Finished current person; exiting step…")
+        _print_game_summary(rec)
+        return 130
+
+    # If game ended, print summary
+    if entry.get("status") != GameStatus.RUNNING:
+        _print_game_summary(rec)
+
+    return 0
+
+
+def cmd_play(interactive: bool, verbose: bool) -> int:
+    bc = BerghainClient()
+    rec = bc.gs.current_game()
+    if not rec:
+        print("No current game. Start one with `berghain new`.", file=sys.stderr)
+        return 1
+    if not rec.policy_path:
+        print("This game has no policy set. Use `berghain set-policy ./policy.py` first.", file=sys.stderr)
+        return 2
+
+    last5: Deque[str] = deque(maxlen=5)
+    renderer = ProgressiveRenderer(rec.constraints.mins, VENUE_CAP, REJECT_CAP)
+    renderer.begin()  # reserve the block before any stdout prints
+
+    _stderr("Autoplaying… (Ctrl-C to stop after finishing the current person)")
+
+    guard = _SoftSigint()
+    guard.arm()
+
+    try:
+        while rec.status == GameStatus.RUNNING and rec.admitted_count < VENUE_CAP:
+            if rec.rejected_count >= 20000:
+                break
+
+            try:
+                entry = bc.step(rec)  # soft SIGINT -> flag set, call completes
+            except PolicyLoadError as e:
+                print(f"Policy error: {e}", file=sys.stderr)
+                return 2
+
+            line = _raw_line(entry, rec.constraints.mins)
+            if verbose:
+                _stderr(line)
+            last5.append(line)
+            renderer.render(last5, rec.tallies,
+                            entry.get("admitted_count", rec.admitted_count),
+                            entry.get("rejected_count", rec.rejected_count))
+
+            if guard.triggered:
+                _stderr("Interrupted. Finished current person; exiting autoplay…")
+                break
+
+            if rec.status != GameStatus.RUNNING:
+                break
+
+            if interactive:
+                # Allow Ctrl-C to abort prompt immediately (no in-flight request)
+                guard.disarm()
+                try:
+                    input("Press Enter for next…")
+                except KeyboardInterrupt:
+                    _stderr("Interrupted at prompt. Exiting after finishing current person…")
+                    break
+                finally:
+                    guard.arm()
+
+    finally:
+        guard.disarm()
+        bc.gs.save()
+
+    _print_game_summary(rec)
+    return 0
+
+
+def cmd_list() -> int:
+    bc = BerghainClient()
+    games = bc.gs.list_games()
+    if not games:
+        print("No games found.")
+        return 0
+    for gid, (status, scenario, admitted, rejected) in games.items():
+        star = "*" if bc.gs.current_game_id == gid else " "
+        print(f"{star} {gid}  scen={scenario}  status={status}  A={admitted}  R={rejected}")
+    return 0
+
+def cmd_switch(idprefix: str) -> int:
+    bc = BerghainClient()
+    ok = bc.gs.set_current_game(idprefix)
+    if not ok:
+        print("No unique match for that id prefix.", file=sys.stderr)
+        return 1
+    bc.gs.save()
+    print(f"Switched current game to: {bc.gs.current_game_id}")
+    return 0
+
+def main(argv: Optional[list[str]] = None) -> int:
+    args = docopt(USAGE, argv=argv, version="berghain 0.0.1")
+    if args["set-player"]:
+        return cmd_set_player(args["<player>"])
+    elif args["new"]:
+        scen = int(args["--scenario"] or 1)
+        if scen not in (1, 2, 3):
+            print("--scenario must be 1, 2, or 3.", file=sys.stderr)
+            return 2
+        return cmd_new_game(scen)
+    elif args["play"]:
+        return cmd_play(bool(args["--interactive"]), bool(args["--verbose"]))
+    elif args["step"]:
+        return cmd_step(bool(args["--verbose"]))
+    elif args["list"]:
+        return cmd_list()
+    elif args["status"]:
+        return cmd_status(args["<idprefix>"])
+    elif args["switch"]:
+        return cmd_switch(args["<idprefix>"])
+    elif args["set-policy"]:
+        if args["default"]:
+            return cmd_set_policy("default")
+        else:
+            return cmd_set_policy(args["<file>"])
+    elif args["check-policy"]:
+        return cmd_check_policy(args["<file>"])
+    else:
+        print("Unknown command", file=sys.stderr)
+        return 1
+
diff --git a/berghain/config.py b/berghain/config.py
new file mode 100644
index 0000000..53a5a34
--- /dev/null
+++ b/berghain/config.py
@@ -0,0 +1,5 @@
+import os
+BASE_URL="https://berghain.challenges.listenlabs.ai"
+STATE_PATH = os.path.expanduser("~/.berghain_state.json")
+VENUE_CAP = 1000
+REJECT_CAP = 20000
diff --git a/berghain/core.py b/berghain/core.py
new file mode 100644
index 0000000..f15a48a
--- /dev/null
+++ b/berghain/core.py
@@ -0,0 +1,246 @@
+from __future__ import annotations
+
+from typing import Any, Dict, Iterator, Optional
+
+from .state import GameState, GameRecord, Constraints, AttributeStats, GameStatus
+from .api import APIClient
+from .policy import load_policy, PolicyLoadError, call_policy, default_policy_path
+from .config import VENUE_CAP, REJECT_CAP
+
+
+class BerghainClient:
+    def __init__(self) -> None:
+        self.gs = GameState.load()
+        self.api = APIClient()
+
+    # ----------------------------
+    # Player / game lifecycle
+    # ----------------------------
+
+    def set_player(self, uuid: str) -> None:
+        """Sets the player ID and persists state."""
+        if not uuid:
+            raise ValueError("Player ID cannot be empty")
+        self.gs.player_id = uuid
+        self.gs.save()
+
+    def new_game(self, scenario: int) -> GameRecord:
+        """
+        Starts a new game (policy is optional and may be attached later).
+        Returns the created GameRecord.
+        """
+        resp = self.api.new_game(self.gs.player_id, scenario)
+
+        prev = self.gs.current_game()
+        inherit_path = getattr(prev, "policy_path", None) if prev else None
+        policy_path = inherit_path or default_policy_path()
+
+        rec = GameRecord(
+            game_id=resp["gameId"],
+            initial_raw=resp,
+            scenario=scenario,
+            status=GameStatus.RUNNING,
+            constraints=Constraints(
+                mins={c["attribute"]: int(c["minCount"]) for c in resp.get("constraints", [])}
+            ),
+            stats=AttributeStats(
+                relative_frequencies=resp.get("attributeStatistics", {}).get("relativeFrequencies", {}) or {},
+                correlations=resp.get("attributeStatistics", {}).get("correlations", {}) or {},
+            ),
+            admitted_count=0,
+            rejected_count=0,
+            next_person=None,
+            tallies={},
+            policy_path=policy_path,
+        )
+        self.gs.add_game(rec)
+        self.gs.save()
+        return rec
+
+    def set_policy_for_current_game(self, policy_path: str) -> GameRecord:
+        """
+        Attach/replace a policy for the current game (validates it first).
+        """
+        rec = self.gs.current_game()
+        if not rec:
+            raise ValueError("No current game to set a policy for.")
+        load_policy(policy_path)  # validate before saving
+        rec.policy_path = policy_path
+        self.gs.save()
+        return rec
+
+    # ----------------------------
+    # Internals
+    # ----------------------------
+
+    def _get_decide_callable(self, rec: GameRecord):
+        """
+        Load and return the user's decide() callable for the given game record.
+        Raises PolicyLoadError if missing/invalid.
+        """
+        if not rec.policy_path:
+            raise PolicyLoadError("No policy set for this game. Use `berghain set-policy <file>`.")
+        return load_policy(rec.policy_path)
+
+    def _update_from_decide_response(self, rec: GameRecord, data: Dict[str, Any]) -> None:
+        """
+        Update local record from /decide-and-next API response.
+        """
+        status = (data.get("status") or "").lower()
+        if status in {GameStatus.RUNNING, GameStatus.COMPLETED, GameStatus.FAILED}:
+            rec.status = status
+        rec.admitted_count = int(data.get("admittedCount", rec.admitted_count))
+        rec.rejected_count = int(data.get("rejectedCount", rec.rejected_count))
+        rec.next_person = data.get("nextPerson")
+        self.gs.save()
+
+    def fetch_first_person(self, rec: GameRecord) -> None:
+        """
+        Fetch the first person (personIndex=0). The API allows 'accept' to be omitted for index 0.
+        """
+        data = self.api.decide_and_next(rec.game_id, person_index=0, accept=None)
+        self._update_from_decide_response(rec, data)
+
+    def _remaining_need(self, rec: GameRecord) -> Dict[str, int]:
+        """
+        Compute remaining shortfall per constrained attribute based on local tallies.
+        """
+        return {
+            a: max(0, rec.constraints.mins.get(a, 0) - rec.tallies.get(a, 0))
+            for a in rec.constraints.mins
+        }
+
+    # ----------------------------
+    # Step / autoplay
+    # ----------------------------
+
+    def step(self, rec: GameRecord) -> Dict[str, Any]:
+        """
+        Execute ONE decision step using the current policy and return a transcript entry:
+
+        {
+          "personIndex": int | None,
+          "accept": bool,
+          "attributes": {attr: bool},
+          "remaining_need_before": {attr: int},
+          "admitted_count": int,   # after API call
+          "rejected_count": int,   # after API call
+          "status": "running"|"completed"|"failed",
+          "reason": str            # human-readable reason (from policy)
+        }
+
+        If no move is possible (e.g., game not running), returns a no-op entry with current counters.
+        Raises PolicyLoadError if no policy is set/invalid.
+        """
+        # If the game isn't running, return a no-op entry.
+        if rec.status != GameStatus.RUNNING:
+            return {
+                "personIndex": None,
+                "accept": False,
+                "attributes": {},
+                "remaining_need_before": self._remaining_need(rec),
+                "admitted_count": rec.admitted_count,
+                "rejected_count": rec.rejected_count,
+                "status": rec.status,
+                "reason": "Game not running",
+            }
+
+        # Ensure we have a person ready.
+        if not rec.next_person:
+            self.fetch_first_person(rec)
+            if not rec.next_person:
+                return {
+                    "personIndex": None,
+                    "accept": False,
+                    "attributes": {},
+                    "remaining_need_before": self._remaining_need(rec),
+                    "admitted_count": rec.admitted_count,
+                    "rejected_count": rec.rejected_count,
+                    "status": rec.status,
+                    "reason": "No next person",
+                }
+
+        # Load user policy decide() and compute decision.
+        decide = self._get_decide_callable(rec)
+
+        idx = int(rec.next_person.get("personIndex", 0))
+        attrs: Dict[str, bool] = rec.next_person.get("attributes", {}) or {}
+        remaining_need_before = self._remaining_need(rec)
+
+        accept = call_policy(
+            decide,
+            attributes=attrs,
+            tallies=rec.tallies,
+            mins=rec.constraints.mins,
+            admitted_count=rec.admitted_count,
+            venue_cap=VENUE_CAP,
+            # New optional context for policies that opt in:
+            p=rec.stats.relative_frequencies,
+            corr=rec.stats.correlations,
+            stats=rec.stats,  # optional: full object if the policy wants it
+        )
+
+        # Ask the policy module for its last explanation
+        reason = None
+        if hasattr(decide, "__globals__"):
+            pol = decide.__globals__
+            if "policy_reason" in pol and callable(pol["policy_reason"]):
+                reason = pol["policy_reason"]()
+
+        # Update local tallies pre-API so they reflect our chosen accept.
+        if accept:
+            rec.bump_tallies(attrs)
+
+        # Send decision, update state.
+        data = self.api.decide_and_next(rec.game_id, person_index=idx, accept=bool(accept))
+        self._update_from_decide_response(rec, data)
+
+        # Build the entry dict (with reason included).
+        entry = {
+            "personIndex": idx,
+            "accept": bool(accept),
+            "attributes": attrs,
+            "remaining_need_before": remaining_need_before,
+            "admitted_count": rec.admitted_count,
+            "rejected_count": rec.rejected_count,
+            "status": rec.status,
+            "reason": reason or ("Admit" if accept else "Reject"),
+        }
+
+        return entry
+
+
+    def autoplay_stream(
+        self,
+        rec: GameRecord,
+        max_steps: Optional[int] = None,
+    ) -> Iterator[Dict[str, Any]]:
+        """
+        Yield transcript entries (same shape as step()) until completion/failure.
+        Stops early if max_steps is provided.
+        """
+        # Ensure policy exists before starting (raises PolicyLoadError if missing/invalid).
+        self._get_decide_callable(rec)
+
+        # Ensure we have the first person available to step on.
+        if rec.next_person is None and rec.status == GameStatus.RUNNING:
+            self.fetch_first_person(rec)
+
+        steps = 0
+        while rec.status == GameStatus.RUNNING:
+            if rec.admitted_count >= VENUE_CAP or rec.rejected_count >= REJECT_CAP:
+                break
+            entry = self.step(rec)
+            yield entry
+            steps += 1
+            if max_steps is not None and steps >= max_steps:
+                break
+
+    def autoplay(self, rec: GameRecord) -> GameRecord:
+        """
+        Run until completion/failure using the current policy (drains the stream).
+        """
+        for _ in self.autoplay_stream(rec):
+            pass
+        return rec
+
diff --git a/berghain/default_policy.py b/berghain/default_policy.py
new file mode 100644
index 0000000..e64da2a
--- /dev/null
+++ b/berghain/default_policy.py
@@ -0,0 +1,78 @@
+# default_policy.py
+# Built-in default policy for Berghain Challenge.
+#
+# Strategy:
+#   - If all minimums are satisfied, admit freely.
+#   - Admit creatives until quota met (they are rare).
+#   - Admit Berlin locals until quota met (they are required in bulk).
+#   - Otherwise admit if candidate helps at least one unmet need,
+#     provided there is still enough capacity to meet all others.
+#   - Reject if candidate does not help with unmet needs
+#     or would make quotas infeasible.
+#
+# The policy records a human-readable reason string for each decision,
+# retrievable via policy_reason(). Your UI can display this directly.
+
+import sys
+
+DEBUG = False
+_last_reason: str | None = None
+
+
+def _set_reason(msg: str) -> None:
+    """Set the last decision reason (and echo to stderr if DEBUG)."""
+    global _last_reason
+    _last_reason = msg
+    if DEBUG:
+        print(msg, file=sys.stderr, flush=True)
+
+
+def policy_reason() -> str | None:
+    """Return the most recent decision reason string."""
+    return _last_reason
+
+
+def decide(attributes: dict[str, bool],
+           tallies: dict[str, int],
+           mins: dict[str, int],
+           admitted_count: int,
+           venue_cap: int,
+           **_) -> bool:
+    """Return True to admit, False to reject."""
+    S = venue_cap - admitted_count
+    need = {a: max(0, mins[a] - tallies.get(a, 0)) for a in mins}
+
+    # 1. All quotas already met
+    if sum(need.values()) == 0:
+        _set_reason("Admit (all minimums satisfied)")
+        return True
+
+    # 2. Admit creatives until quota met
+    if need.get("creative", 0) > 0 and attributes.get("creative", False):
+        _set_reason("Admit creative (rare + still needed)")
+        return True
+
+    # 3. Admit locals until quota met
+    if need.get("berlin_local", 0) > 0 and attributes.get("berlin_local", False):
+        _set_reason("Admit berlin_local (large quota still unmet)")
+        return True
+
+    # 4. Admit if candidate helps an unmet need and feasibility is safe
+    helps = [a for a in need if need[a] > 0 and attributes.get(a, False)]
+    if helps:
+        S_after = S - 1
+        remaining_after = sum(
+            max(0, need[a] - (1 if attributes.get(a, False) else 0))
+            for a in need
+        )
+        if S_after >= remaining_after:
+            _set_reason(f"Admit (helps {', '.join(helps)}, feasibility ok)")
+            return True
+        else:
+            _set_reason("Reject (would break feasibility)")
+            return False
+
+    # 5. Reject if they don’t help
+    _set_reason("Reject (no help to unmet needs)")
+    return False
+
diff --git a/berghain/policy.py b/berghain/policy.py
new file mode 100644
index 0000000..86eabd5
--- /dev/null
+++ b/berghain/policy.py
@@ -0,0 +1,50 @@
+# policy.py (loader/utilities for user-supplied policy scripts)
+from __future__ import annotations
+
+import importlib.util
+import inspect
+import os
+from types import ModuleType
+from typing import Callable, Any
+
+
+class PolicyLoadError(Exception):
+    pass
+
+
+def _import_module_from_path(path: str) -> ModuleType:
+    if not os.path.exists(path):
+        raise PolicyLoadError(f"Policy file not found: {path}")
+    spec = importlib.util.spec_from_file_location("user_policy", path)
+    if spec is None or spec.loader is None:
+        raise PolicyLoadError("Could not load policy module spec.")
+    mod = importlib.util.module_from_spec(spec)
+    try:
+        spec.loader.exec_module(mod)  # type: ignore[reportAttributeAccessIssue]
+    except Exception as e:
+        raise PolicyLoadError(f"Error executing policy module: {e}") from e
+    return mod
+
+def default_policy_path() -> str:
+    return os.path.join(os.path.dirname(__file__), "default_policy.py")
+
+def load_policy(path: str) -> Callable[..., bool]:
+    """
+    Load a user policy file and return the `decide` callable.
+    """
+    mod = _import_module_from_path(path)
+    decide = getattr(mod, "decide", None)
+    if not callable(decide):
+        raise PolicyLoadError("Policy script must define a callable `decide(...)`.")
+    return decide
+
+
+def call_policy(decide: Callable[..., Any], **kwargs: Any) -> Any:
+    """
+    Call `decide` with only the parameters it accepts.
+    This lets us pass new optional context (like p, corr) without breaking old policies.
+    """
+    sig = inspect.signature(decide)
+    accepted = {k: v for k, v in kwargs.items() if k in sig.parameters}
+    return decide(**accepted)
+
diff --git a/berghain/state.py b/berghain/state.py
new file mode 100644
index 0000000..4d5f583
--- /dev/null
+++ b/berghain/state.py
@@ -0,0 +1,124 @@
+from __future__ import annotations
+
+from dataclasses import dataclass, field, asdict
+from typing import Any, Dict, Tuple
+import json
+import os
+
+from .config import STATE_PATH
+
+
+class GameStatus(str):
+    RUNNING = "running"
+    COMPLETED = "completed"
+    FAILED = "failed"
+
+
+@dataclass
+class Constraints:
+    mins: Dict[str, int] = field(default_factory=dict)
+
+
+@dataclass
+class AttributeStats:
+    relative_frequencies: Dict[str, float] = field(default_factory=dict)
+    correlations: Dict[str, Dict[str, float]] = field(default_factory=dict)
+
+
+@dataclass
+class GameRecord:
+    game_id: str
+    initial_raw: Dict[str, Any] | None = None
+    scenario: int | None = None
+    status: str = GameStatus.RUNNING
+    constraints: Constraints = field(default_factory=Constraints)
+    stats: AttributeStats = field(default_factory=AttributeStats)
+    admitted_count: int = 0
+    rejected_count: int = 0
+    next_person: Dict[str, Any] | None = None
+    tallies: Dict[str, int] = field(default_factory=dict)
+    policy_path: str | None = None  # <— path to the user-provided policy script
+
+    def bump_tallies(self, attributes: Dict[str, bool]) -> None:
+        for k, v in attributes.items():
+            if v:
+                self.tallies[k] = self.tallies.get(k, 0) + 1
+
+    def raw(self) -> Dict[str, Any]:
+        return asdict(self)
+
+    @classmethod
+    def from_dict(cls, d: Dict[str, Any]) -> "GameRecord":
+        constraints = Constraints(**(d.get("constraints") or {}))
+        stats = AttributeStats(**(d.get("stats") or {}))
+        return cls(
+            game_id=d["game_id"],
+            initial_raw=d.get("initial_raw"),
+            scenario=d.get("scenario"),
+            status=d.get("status", GameStatus.RUNNING),
+            constraints=constraints,
+            stats=stats,
+            admitted_count=d.get("admitted_count", 0),
+            rejected_count=d.get("rejected_count", 0),
+            next_person=d.get("next_person"),
+            tallies=d.get("tallies", {}) or {},
+            policy_path=d.get("policy_path"),
+        )
+
+
+@dataclass
+class GameState:
+    player_id: str | None = None
+    current_game_id: str | None = None
+    games: Dict[str, GameRecord] = field(default_factory=dict)
+
+    def save(self) -> None:
+        payload = {
+            "player_id": self.player_id,
+            "current_game_id": self.current_game_id,
+            "games": {gid: g.raw() for gid, g in self.games.items()},
+        }
+        os.makedirs(os.path.dirname(STATE_PATH), exist_ok=True)
+        with open(STATE_PATH, "w", encoding="utf-8") as f:
+            json.dump(payload, f, indent=2)
+
+    @classmethod
+    def load(cls) -> "GameState":
+        try:
+            with open(STATE_PATH, "r", encoding="utf-8") as f:
+                raw = json.load(f)
+        except FileNotFoundError:
+            return cls()
+        gs = cls()
+        gs.player_id = raw.get("player_id")
+        gs.current_game_id = raw.get("current_game_id")
+        for gid, graw in (raw.get("games") or {}).items():
+            gs.games[gid] = GameRecord.from_dict(graw)
+        return gs
+
+    def add_game(self, rec: GameRecord) -> None:
+        self.games[rec.game_id] = rec
+        self.current_game_id = rec.game_id
+
+    def current_game(self) -> GameRecord | None:
+        return self.games.get(self.current_game_id) if self.current_game_id else None
+
+    def set_current_game(self, game_id_prefix: str) -> bool:
+        g = self.get_game(game_id_prefix)
+        if g:
+            self.current_game_id = g.game_id
+            return True
+        return False
+
+    def get_game(self, game_id_prefix: str) -> GameRecord | None:
+        matches = [gid for gid in self.games if gid.startswith(game_id_prefix)]
+        if len(matches) == 1:
+            return self.games[matches[0]]
+        return None
+
+    def list_games(self) -> Dict[str, Tuple[str, int | None, int, int]]:
+        return {
+            gid: (g.status, g.scenario, g.admitted_count, g.rejected_count)
+            for gid, g in self.games.items()
+        }
+
diff --git a/berghain/ui.py b/berghain/ui.py
new file mode 100644
index 0000000..8d102aa
--- /dev/null
+++ b/berghain/ui.py
@@ -0,0 +1,181 @@
+from __future__ import annotations
+
+import shutil
+from typing import Deque, Dict, List
+
+# ANSI
+RESET = "\x1b[0m"
+BOLD = "\x1b[1m"
+DIM = "\x1b[2m"
+FG_RED = "\x1b[31m"
+FG_GREEN = "\x1b[32m"
+FG_CYAN = "\x1b[36m"
+FG_YELLOW = "\x1b[33m"
+ERASE_EOL = "\x1b[K"
+
+def _term_width() -> int:
+    return shutil.get_terminal_size(fallback=(80, 20)).columns
+
+def _bar(current: int, target: int, width: int, color: str = "", label: str = "", pad_to: int | None = None) -> str:
+    """
+    Draw a bar and show " current/target" with current optionally right-aligned to pad_to digits.
+    """
+    target = max(1, target)
+    ratio = min(1.0, max(0.0, current / target))
+    filled = int(ratio * width)
+    bar = "[" + "#" * filled + "-" * (width - filled) + "]"
+    label_part = f" {label}" if label else ""
+    # right-align current to pad_to (defaults to len(str(target)))
+    w = pad_to if pad_to is not None else len(str(target))
+    cur_s = str(current).rjust(w)
+    val = f"{cur_s}/{target}"
+    return f"{color}{bar}{RESET}{label_part} {DIM}{val}{RESET}"
+
+class ProgressiveRenderer:
+    """
+    Flicker-free renderer that:
+      - Prints the whole block ONCE (with real newlines) and leaves the cursor
+        on the LAST line of the block (no extra newline).
+      - Updates by moving the cursor up to the top of the block, overwriting
+        ONLY changed lines, then returning to the bottom line (never beyond).
+      - The initial scaffold reserves blank lines for the constraints to avoid
+        label vs. label+bar duplication on some terminals.
+    """
+
+    def __init__(self, mins: Dict[str, int], venue_cap: int, reject_cap: int) -> None:
+        self.mins = dict(mins)
+        self.labels = sorted(self.mins.keys())
+        self.venue_cap = venue_cap
+        self.reject_cap = reject_cap
+
+        # Layout indices within the reserved block
+        self.recent_count = 5
+        self.recent_title = 0
+        self.recent_lines_start = 1              # 5 lines: [1..5]
+        self.recent_lines_end = self.recent_lines_start + self.recent_count - 1  # 5
+
+        self.constraints_header = self.recent_lines_end + 1 + 1   # +1 spacer
+        self.constraints_start = self.constraints_header + 1
+        self.constraints_lines = len(self.labels)
+        self.constraints_end = self.constraints_start + self.constraints_lines - 1
+
+        self.totals_header = self.constraints_end + 1 + 1
+        self.total_admit_line = self.totals_header + 1
+        self.total_reject_line = self.totals_header + 2
+
+        self.total_lines = self.total_reject_line + 1  # number of lines in block
+
+        self._cache: Dict[int, str] = {}  # last rendered content per line
+
+        term = _term_width()
+        self.bar_inner_width = max(10, min(50, term - 30))
+
+        # Keep the cursor "inside" the block; we park it on the last line.
+        self._cur_line = self.total_lines - 1
+        self._begun = False
+
+    # ---- cursor movement helpers (relative to *current* position) ----
+
+    def _move_to_top(self) -> None:
+        if self._cur_line > 0:
+            print(f"\x1b[{self._cur_line}F", end="")  # up to BOL of top line
+            self._cur_line = 0
+        else:
+            print("\r", end="")
+
+    def _move_to_line(self, target: int) -> None:
+        delta = target - self._cur_line
+        if delta > 0:
+            print(f"\x1b[{delta}E", end="")  # down to BOL of target
+        elif delta < 0:
+            print(f"\x1b[{-delta}F", end="")  # up to BOL of target
+        else:
+            print("\r", end="")
+        self._cur_line = target
+
+    def _write_line(self, text: str) -> None:
+        print("\r" + text + ERASE_EOL, end="")
+
+    def _return_to_bottom(self) -> None:
+        if self._cur_line < self.total_lines - 1:
+            print(f"\x1b[{(self.total_lines - 1) - self._cur_line}E", end="")
+            self._cur_line = self.total_lines - 1
+        print("\r", end="", flush=True)
+
+    # ---- public API ----
+
+    def begin(self) -> None:
+        if self._begun:
+            return
+        # Print the static block once. IMPORTANT: the constraint lines are BLANK here.
+        lines: List[str] = []
+        # Recent
+        lines.append(f"{BOLD}Recent decisions{RESET}")
+        for _ in range(self.recent_count):
+            lines.append("")
+        # spacer
+        lines.append("")
+        # Constraints header + blank lines (reserved)
+        lines.append(f"{BOLD}Constraints{RESET}")
+        for _ in self.labels:
+            lines.append("")  # reserve a blank line for each constraint
+        # spacer
+        lines.append("")
+        # Totals
+        lines.append(f"{BOLD}Totals{RESET}")
+        lines.append("")  # admitted bar
+        lines.append("")  # rejected count
+
+        for i, txt in enumerate(lines):
+            end = "" if i == len(lines) - 1 else "\n"
+            print(txt, end=end)
+            self._cache[i] = txt
+
+        self._cur_line = self.total_lines - 1
+        print("\r", end="", flush=True)
+        self._begun = True
+
+    def render(self, recent: Deque[str], tallies: Dict[str, int],
+               admitted: int, rejected: int) -> None:
+        """
+        Update the block with current data, overwriting only changed lines.
+        Leaves the cursor on the last line of the block (BOL).
+        """
+        desired: Dict[int, str] = {}
+
+        # Recent (bottom-aligned) — we now expect `recent` to contain
+        # human-readable policy reasons, not raw attribute dumps.
+        pad = max(0, self.recent_count - len(recent))
+        recent_padded = [""] * pad + list(recent)
+        desired[self.recent_title] = f"{BOLD}Recent decisions{RESET}"
+        for i in range(self.recent_count):
+            desired[self.recent_lines_start + i] = recent_padded[i]
+
+        # Constraints (full line: label + bar)
+        desired[self.constraints_header] = f"{BOLD}Constraints{RESET}"
+        for i, a in enumerate(self.labels):
+            have = tallies.get(a, 0)
+            need = self.mins[a]
+            color = FG_CYAN if have < need else FG_YELLOW
+            bar = _bar(have, need, self.bar_inner_width, color, pad_to=len(str(need)))
+            desired[self.constraints_start + i] = f"{a:>16}: {bar}"
+
+        # Totals (right-aligned numbers)
+        desired[self.totals_header] = f"{BOLD}Totals{RESET}"
+        desired[self.total_admit_line] = _bar(
+            admitted, self.venue_cap, self.bar_inner_width,
+            FG_GREEN, "admitted", pad_to=len(str(self.venue_cap))
+        )
+        rej_w = len(str(self.reject_cap))
+        desired[self.total_reject_line] = f"{FG_RED}rejected{RESET} {DIM}{str(rejected).rjust(rej_w)}{RESET}"
+
+        # Minimal redraw
+        self._move_to_top()
+        for idx in range(self.total_lines):
+            new_text = desired.get(idx, "")
+            if self._cache.get(idx) != new_text:
+                self._move_to_line(idx)
+                self._write_line(new_text)
+                self._cache[idx] = new_text
+        self._return_to_bottom()
+
diff --git a/pyproject.toml b/pyproject.toml
new file mode 100644
index 0000000..9f844f3
--- /dev/null
+++ b/pyproject.toml
@@ -0,0 +1,11 @@
+[project]
+name = "berghain"
+version = "0.1.1"
+description = "berghain challenge client"
+dependencies = [
+  "docopt>=0.6",
+  "requests>=2.31",
+]
+
+[project.scripts]
+berghain = "berghain.cli:main"