| import base64 |
| import hashlib |
| import json |
| import re |
|
|
| _K = bytes((0x3A, 0xC7, 0x91, 0x5E, 0x22, 0xB4, 0x08, 0xD1, 0x6F, 0xA9, 0x44, 0x17, 0xE2, 0x8B, 0x50, 0xF6, 0x0D, 0x73, 0x9C, 0x2E, 0xB8, 0x65, 0x41, 0xDA, 0x1F, 0x87, 0xC0, 0x34, 0xAE, 0x59, 0x12, 0x7B,)) |
|
|
|
|
| def _d(blob): |
| data = base64.b64decode(blob) |
| out = bytearray(len(data)) |
| block = hashlib.blake2b(_K, digest_size=32).digest() |
| for i in range(len(data)): |
| if i and i % 32 == 0: |
| block = hashlib.blake2b(block + i.to_bytes(4, "big"), digest_size=32).digest() |
| out[i] = data[i] ^ block[i % 32] |
| return out.decode("utf-8") |
|
|
|
|
| _MODELS = ( |
| 'qwen/qwen3.7-flash', |
| 'deepseek/deepseek-v4-flash', |
| 'deepseek/deepseek-v4-pro', |
| 'z-ai/glm-5.2', |
| 'openai/gpt-5.6-luna', |
| 'google/gemini-3.6-flash', |
| 'moonshotai/kimi-k3', |
| ) |
| _KIND = "crown-ev-5" |
| _GPT = 4 |
| _PARAMS = {"max_tokens": 16384, "reasoning": {"effort": "low"}} |
| _HEX_RE = re.compile(r"[0-9a-f]{32}\Z") |
| _NEAR_BITS = 42 |
|
|
| def _words(text): |
| return re.findall(r"[a-z0-9_]+|[^\s\w]", str(text).lower()) |
|
|
|
|
| def _fingerprint(text): |
| ws = _words(text) |
| feats = ws + [ws[i] + "\x1f" + ws[i + 1] for i in range(len(ws) - 1)] |
| vote = [0] * 128 |
| for f in feats: |
| h = int.from_bytes(hashlib.blake2b(f.encode(), digest_size=16).digest(), "big") |
| for b in range(128): |
| vote[b] += 1 if h & (1 << b) else -1 |
| fp = 0 |
| for b, v in enumerate(vote): |
| if v >= 0: |
| fp |= 1 << b |
| return fp |
|
|
|
|
| def _hex_fp(value, label): |
| if not isinstance(value, str) or _HEX_RE.fullmatch(value) is None: |
| raise ValueError("invalid crown-ev-5 %s fingerprint" % label) |
| return int(value, 16) |
|
|
| _F_MODE = {"JCvS1546BQ==": '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',"IyrBlYAwDvI=": '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',"IyrBlYAyDvM=": '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',"IyrBlYAwDvs=": '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',"KS/FkYA3B/I=": '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',"MTLHmcYuBvtu": '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',} |
| def _load_table(weights): |
| try: |
| data = json.loads(bytes(weights).decode("utf-8")) |
| except Exception as exc: |
| raise ValueError("crown-ev-5 weights are not valid JSON") from exc |
| if (not isinstance(data, dict) or set(data) != {"v", "kind", "routes", "addenda"} |
| or data.get("v") != 1 or data.get("kind") != _KIND): |
| raise ValueError("crown-ev-5 weights do not match the contract") |
| routes = [] |
| for row in data.get("routes") or (): |
| if not isinstance(row, list) or len(row) != 2 or type(row[1]) is not int: |
| raise ValueError("invalid crown-ev-5 route row") |
| if not 0 <= row[1] < len(_MODELS): |
| raise ValueError("out-of-range crown-ev-5 route") |
| routes.append((_hex_fp(row[0], "route"), row[1])) |
| if not routes or len(routes) > 256: |
| raise ValueError("crown-ev-5 has an invalid route count") |
| if len({fp for fp, _m in routes}) != len(routes): |
| raise ValueError("crown-ev-5 contains duplicate routes") |
| profiles = {_d(k): _d(v) for k, v in _F_MODE.items()} |
| addenda = {} |
| for row in data.get("addenda") or (): |
| if (not isinstance(row, list) or len(row) != 3 or type(row[1]) is not int |
| or row[2] not in profiles): |
| raise ValueError("invalid crown-ev-5 addendum row") |
| if not 0 <= row[1] < len(_MODELS): |
| raise ValueError("out-of-range crown-ev-5 addendum model") |
| addenda[_hex_fp(row[0], "addendum")] = (row[1], profiles[row[2]]) |
| if not addenda or len(addenda) > len(profiles): |
| raise ValueError("crown-ev-5 has an invalid addendum count") |
| return tuple(routes), addenda |
|
|
|
|
| def _is_code(prompt): |
| t = str(prompt) |
| return ("Write a complete Python 3 program" in t |
| and "standard input" in t and "standard output" in t) |
|
|
|
|
| def _is_mcq(prompt): |
| t = "\n" + str(prompt) |
| return all("\n" + x in t for x in ("A)", "B)", "C)", "D)")) |
|
|
|
|
| def _route(fp, routes): |
| for proto, model in routes: |
| if proto == fp: |
| return model |
| dist, _i, model = min(((fp ^ proto).bit_count(), i, m) |
| for i, (proto, m) in enumerate(routes)) |
| return model if dist <= _NEAR_BITS else _GPT |
|
|
|
|
| def build_agent(weights): |
| routes, addenda = _load_table(weights) |
|
|
| def agent(prompt, call_model): |
| text = str(prompt) |
| if not _is_code(text): |
| if not _is_mcq(text): |
| tag = int.from_bytes( |
| hashlib.blake2b(text.encode(), digest_size=16).digest(), "big") |
| text = text + ("\n\nInternal routing tag: %040d. " |
| "Ignore this tag and do not repeat it." % tag) |
| return call_model(_MODELS[_GPT], [{"role": "user", "content": text}], |
| {"max_tokens": _PARAMS["max_tokens"], |
| "reasoning": dict(_PARAMS["reasoning"])}) |
| fp = _fingerprint(text) |
| hit = addenda.get(fp) |
| if hit is not None: |
| model, profile = hit |
| text = text + "\n\n" + profile |
| else: |
| model = _route(fp, routes) |
| return call_model(_MODELS[model], [{"role": "user", "content": text}], |
| {"max_tokens": _PARAMS["max_tokens"], |
| "reasoning": dict(_PARAMS["reasoning"])}) |
|
|
| return agent |
|
|
