# { "Depends": "py-genlayer:1jb45aa8ynh2a9c9xn3b7qqh8sm5q93hwfp7jqmwsfhh8jpz09h6" } """ NewsOracle — a GenLayer Intelligent Contract that produces multi-validator AI consensus analyses of crypto / Web3 news articles. Each `analyze()` call: 1. Has every validator independently run an LLM over the same article. 2. Reaches consensus via the equivalence principle (sentiment label and category must match exactly; the score must be within a tolerance). 3. Persists the consensus verdict on-chain forever, keyed by article id. """ from genlayer import * import json import hashlib import typing SYSTEM_PROMPT = """You are a professional Web3 / crypto market analyst with deep expertise in blockchain, DeFi, NFTs, and crypto markets. Analyse the provided news article and respond with ONLY a valid JSON object — no markdown, no explanation, no preamble: { "sentiment": "bullish" | "bearish" | "neutral", "sentiment_score": , "summary": "<2-3 sentence plain English summary of what happened and why it matters>", "key_points": ["", "", "", ""], "market_impact": "<1-2 sentences on potential short-term market impact>", "risk_level": "low" | "medium" | "high", "category": "DeFi" | "NFT" | "Layer1" | "Layer2" | "Regulation" | "Macro" | "Exchange" | "AI" | "Infrastructure" | "Other", "entities": ["", ...], "tldr": "" }""" # Validators reach consensus when these conditions are jointly satisfied. # Loose enough to absorb LLM variance, tight enough that bad-faith leaders # can't slip distorted analyses past honest validators. EQUIVALENCE_PRINCIPLE = ( "`sentiment`, `risk_level`, and `category` fields MUST match exactly. " "`sentiment_score` values must be within 20 points of each other. " "`key_points` and `entities` lists must cover substantially the same " "themes / names (order and wording may differ). " "`summary`, `market_impact`, and `tldr` may be phrased differently but " "must convey the same overall conclusion." ) def _article_id(title: str, content: str) -> str: """Stable id from the article text — lets clients dedupe cheaply.""" h = hashlib.sha256() h.update(title.strip().encode("utf-8")) h.update(b"\x00") h.update(content.strip().encode("utf-8")) return h.hexdigest()[:24] class NewsOracle(gl.Contract): # article_id -> JSON-encoded consensus analysis analyses: TreeMap[str, str] # rolling log of analyzed article ids, newest last history: DynArray[str] def __init__(self): pass # -- Reads -------------------------------------------------------------- @gl.public.view def get_analysis(self, article_id: str) -> str: if article_id in self.analyses: return self.analyses[article_id] return "" @gl.public.view def has_analysis(self, title: str, content: str) -> bool: return _article_id(title, content) in self.analyses @gl.public.view def get_recent(self) -> list[str]: items = list(self.history) items.reverse() return items[:50] # -- Writes ------------------------------------------------------------- @gl.public.write def analyze(self, title: str, content: str) -> typing.Any: """ Reach multi-validator consensus on the analysis of a single article and persist the verdict on-chain. """ if not title: raise Exception("title is required") article_id = _article_id(title, content) # Idempotent: serve cached consensus if we've already analysed this. if article_id in self.analyses: cached = json.loads(self.analyses[article_id]) cached["cached"] = True return cached # Trim aggressively — every byte the leader returns has to be agreed # upon and stored on-chain. body = (content or title)[:3000] text = f"Title: {title}\n\nContent: {body}" def leader_and_validator() -> str: # Each validator independently asks its own LLM the same question. prompt = f"{SYSTEM_PROMPT}\n\n=== Article ===\n{text}\n=== End ===" raw = gl.nondet.exec_prompt(prompt) cleaned = raw.replace("```json", "").replace("```", "").strip() # Tolerate a model that wraps JSON in extra prose. start = cleaned.find("{") end = cleaned.rfind("}") if start == -1 or end == -1 or end <= start: raise Exception(f"LLM did not return JSON: {raw[:200]}") parsed = json.loads(cleaned[start : end + 1]) # Schema validation — reject obvious garbage early so a bad # leader can't poison state. for field in ( "sentiment", "sentiment_score", "summary", "key_points", "market_impact", "risk_level", "category", "entities", "tldr", ): if field not in parsed: raise Exception(f"missing field in LLM response: {field}") score = int(parsed["sentiment_score"]) if score < 0 or score > 100: raise Exception(f"sentiment_score out of range: {score}") parsed["sentiment_score"] = score parsed["sentiment"] = str(parsed["sentiment"]).lower() parsed["risk_level"] = str(parsed["risk_level"]).lower() # Canonical ordering so validator string-comparison is stable. return json.dumps(parsed, sort_keys=True) verdict_json = gl.eq_principle.prompt_comparative( leader_and_validator, principle=EQUIVALENCE_PRINCIPLE, ) verdict = json.loads(verdict_json) verdict["article_id"] = article_id verdict["title"] = title verdict["cached"] = False stored = json.dumps(verdict, sort_keys=True) self.analyses[article_id] = stored self.history.append(article_id) return verdict

Hash	Type	Status		From	To	Method	GenVM Result	Consensus Result	Time
0xb4ee577b...791daca2	Call	FINALIZED	IN	0xDcb56A...290ed0	0x713300...1826ae	analyze	SUCCESS	Accepted	3 months ago
0x200db780...9689ec8d	Deploy	FINALIZED	IN	0xDcb56A...290ed0	0x713300...1826ae	(constructor)	SUCCESS	Accepted	3 months ago