# v0.2.17 # { "Depends": "py-genlayer:1jb45aa8ynh2a9c9xn3b7qqh8sm5q93hwfp7jqmwsfhh8jpz09h6" } import json import re from dataclasses import dataclass from datetime import datetime, timezone from genlayer import * # ---------------------------------------------------------------------------- # Payment architecture: GenLayer is the adjudication layer ONLY. It never # escrows or moves real value. Prize money is real USDC held in a dedicated # escrow contract on Base Sepolia (contracts/base/MemeOlympicsEscrow.sol). # This contract just judges memes and produces a deterministic winners list # with USDC amounts owed; a backend relayer reads that list once a # competition finalizes and calls the escrow contract on Base Sepolia to # make the payout claimable there. See `mark_prizes_relayed` below — the # relayer calls it (idempotency marker only, moves no value here) after its # Base Sepolia transaction confirms. # ---------------------------------------------------------------------------- # ---------------------------------------------------------------------------- # Error classification prefixes. # # Validators use these to decide how to compare failures: # EXPECTED — deterministic business-logic failures; exact match required # EXTERNAL — external API 4xx-style failures; exact match required # TRANSIENT — network/5xx flakiness; validators agree if both are transient # LLM — LLM misbehaviour; validators always disagree, forcing leader # rotation instead of locking bad state on-chain # ---------------------------------------------------------------------------- ERROR_EXPECTED = "[EXPECTED]" ERROR_EXTERNAL = "[EXTERNAL]" ERROR_TRANSIENT = "[TRANSIENT]" ERROR_LLM = "[LLM_ERROR]" # ---------------------------------------------------------------------------- # Domain constants — statuses stored as plain strings (Enums are not # persistable in GenLayer storage). # ---------------------------------------------------------------------------- COMP_STATUS_CREATED = "created" COMP_STATUS_OPEN = "open" COMP_STATUS_JUDGING = "judging" COMP_STATUS_FINALIZED = "finalized" COMP_STATUS_CANCELLED = "cancelled" SUB_STATUS_PENDING = "pending" SUB_STATUS_EVALUATED = "evaluated" SUB_STATUS_REJECTED = "rejected" SUB_STATUS_DISQUALIFIED = "disqualified" SUB_STATUS_WINNER = "winner" DISPUTE_STATUS_OPEN = "open" DISPUTE_STATUS_UPHELD = "upheld" DISPUTE_STATUS_REJECTED = "rejected" PLAGIARISM_ORIGINAL = "original" PLAGIARISM_SUSPICIOUS = "suspicious" PLAGIARISM_COPIED = "copied" # The nine judging criteria, in canonical order. Weights are configurable # on-chain (basis points, sum == 10000) but the criteria set is fixed so # leader and validators always score the same dimensions. CRITERIA = [ "originality", "humor", "relevance", "timing", "irony", "cultural_awareness", "crypto_native_understanding", "contextual_intelligence", "creativity", ] # Default criterion weights in basis points (sum = 10000). DEFAULT_WEIGHTS_BP = { "originality": 1600, "humor": 1600, "relevance": 1200, "timing": 900, "irony": 800, "cultural_awareness": 1000, "crypto_native_understanding": 1400, "contextual_intelligence": 700, "creativity": 800, } # Consensus tolerances — tight enough that two honest but slightly-differing # LLM runs still agree on genuine outliers, while dishonest/broken leaders # still get caught. (Tightened from 15/5/7-of-9 after review: the previous # tolerances let a leader and validator disagree on nearly 1/3 of the total # score band and still reach consensus.) SCORE_TOLERANCE = 10 # total score points out of 100 CRITERION_TOLERANCE = 3 # per-criterion points out of 10 CONFIDENCE_TOLERANCE = 25 # plagiarism confidence, out of 100 CRITERIA_AGREEMENT_MIN = 8 # of 9 criteria must be within tolerance # Hard limits (anti-gaming / compute-bound safety). MAX_TITLE_LEN = 120 MAX_CAPTION_LEN = 600 MAX_URL_LEN = 500 MAX_TAGS = 8 MAX_TAG_LEN = 32 MAX_SUBMISSIONS_PER_COMPETITION = 500 MAX_DISPUTE_REASON_LEN = 1000 USDC_DECIMALS = 6 # prize amounts are stored in USDC's own base units (not atto) # Winners cannot become payable until this window has elapsed. It is a # protocol rule, not a scheduler convention: `mark_prizes_relayed` enforces # it and the Base escrow also refuses early claims. CHALLENGE_PERIOD_SECONDS = 24 * 60 * 60 # ---------------------------------------------------------------------------- # Storage dataclasses. # # Only str / u256 / bool fields are used inside dataclasses; any list-like # or nested structure is serialized to a JSON string. This keeps the storage # layout simple, upgrade-safe (append-only), and avoids nested-container # pitfalls. New fields must only ever be APPENDED at the end. # ---------------------------------------------------------------------------- @allow_storage @dataclass class Competition: id: str title: str theme: str status: str starts_at: str # ISO-8601 string ends_at: str # ISO-8601 string prize_pool_atto: u256 # DEPRECATED, kept 0 — see prize_pool_usdc winner_count: u256 submission_count: u256 created_by: str # hex address of creator created_at: str finalized_at: str winners_json: str # JSON: [{submission_id, rank, reward_usdc}] # Appended: real prize money now lives in the Base Sepolia USDC escrow # (contracts/base/MemeOlympicsEscrow.sol), not here. This is the # declared/informational pool in USDC base units (6 decimals) that # should match what's actually deposited under this competition_id on # that escrow contract — the escrow's own balance is the source of # truth for real funds, this is display/bookkeeping only. prize_pool_usdc: u256 relay_tx_hash: str # Base Sepolia tx hash once winners were relayed # Appended. Unix seconds, derived from the deterministic GenVM transaction # clock at finalization. Payout relay is forbidden before this moment. challenge_ends_at: u256 @allow_storage @dataclass class Submission: id: str competition_id: str author: str # hex address of submitting wallet title: str caption: str image_url: str context_url: str # optional supporting link ("why this is timely") tags_json: str # JSON array of strings status: str submitted_at: str # Evaluation results (zero/empty until evaluated) total_score: u256 # 0..100 criteria_json: str # JSON: {criterion: 0..10} plagiarism_verdict: str # original | suspicious | copied | "" plagiarism_confidence: u256 # 0..100 image_accessible: bool evaluation_summary: str evaluated_at: str # Appended: hard eligibility gate result (see CRITERIA / evaluate_submission). eligible: bool eligibility_reason: str @allow_storage @dataclass class Dispute: id: str submission_id: str competition_id: str opened_by: str reason: str evidence_url: str status: str verdict_summary: str opened_at: str resolved_at: str @allow_storage @dataclass class UserStats: address: str submissions: u256 wins: u256 total_score_accum: u256 rewards_atto: u256 disqualifications: u256 last_submission_at: str class MemeOlympics(gl.Contract): """Weekly AI-judged crypto meme competitions with validator consensus.""" # ------------------------------------------------------------------ # Storage layout — append-only. Never reorder or insert fields. # ------------------------------------------------------------------ owner: Address admins: TreeMap[str, bool] paused: bool # Configuration weights_json: str # JSON {criterion: weight_bp} default_winner_count: u256 min_seconds_between_submissions: u256 # advisory, enforced off-chain max_submissions_per_user: u256 # per competition # Competitions competitions: TreeMap[str, Competition] competition_ids: DynArray[str] active_competition_id: str # Submissions submissions: TreeMap[str, Submission] submission_ids: DynArray[str] competition_submissions_json: TreeMap[str, str] # comp_id -> JSON [sub_id] user_comp_submission_count: TreeMap[str, u256] # f"{comp_id}:{addr}" -> n image_url_seen: TreeMap[str, str] # normalized url -> sub_id # Disputes disputes: TreeMap[str, Dispute] dispute_ids: DynArray[str] submission_dispute_count: TreeMap[str, u256] # Rewards & users — NO real value is ever held here (see payment # architecture note at the top of the file). reward_balances_atto is a # declared/informational USDC-base-units ledger (field name kept for # storage-layout stability), populated at finalize_competition and only # meaningful once the matching competition's prize was actually # deposited on the Base Sepolia escrow contract. user_stats: TreeMap[str, UserStats] reward_balances_atto: TreeMap[str, u256] # declared USDC owed, per address total_rewards_distributed_atto: u256 # declared USDC allocated at finalize total_rewards_claimed_atto: u256 # declared USDC across mark_prizes_relayed calls # Counters / audit total_competitions: u256 total_submissions: u256 total_evaluations: u256 total_disputes: u256 audit_log: DynArray[str] # JSON event strings, capped reads # Evidence for the exceptional INVALID_IMAGE terminal path. Append-only # storage: each entry is the accepted-but-execution-failed evaluation tx # hash observed by the operator. Three distinct commitments are required # before an admin can terminally disqualify a pending entry. unjudgeable_attempt_count: TreeMap[str, u256] unjudgeable_attempt_hashes_json: TreeMap[str, str] # ================================================================== # Constructor # ================================================================== def __init__(self, owner_address: str = ""): # Some deployment paths (RPC vs Studio UI) present a zero sender to # the constructor, so an explicit owner may be passed; it falls back # to the deploying sender when omitted. if owner_address: self.owner = Address(owner_address) else: self.owner = gl.message.sender_address self.paused = False self.admins[self._addr_str(self.owner)] = True self.weights_json = json.dumps(DEFAULT_WEIGHTS_BP, sort_keys=True) self.default_winner_count = u256(3) self.min_seconds_between_submissions = u256(60) self.max_submissions_per_user = u256(3) self.active_competition_id = "" self.total_rewards_distributed_atto = u256(0) self.total_rewards_claimed_atto = u256(0) self.total_competitions = u256(0) self.total_submissions = u256(0) self.total_evaluations = u256(0) self.total_disputes = u256(0) self._log("contract_deployed", {"owner": self._addr_str(self.owner)}) # ================================================================== # Internal helpers — deterministic only # ================================================================== @staticmethod def _addr_str(addr: Address) -> str: return addr.as_hex def _sender(self) -> str: return self._addr_str(gl.message.sender_address) def _require_not_paused(self) -> None: if self.paused: raise gl.vm.UserError(f"{ERROR_EXPECTED} Contract is paused") def _require_owner(self) -> None: if gl.message.sender_address != self.owner: raise gl.vm.UserError(f"{ERROR_EXPECTED} Only owner may call this") def _require_admin(self) -> None: sender = self._sender() if not (sender in self.admins and self.admins[sender]): raise gl.vm.UserError(f"{ERROR_EXPECTED} Only admin may call this") def _require_creator_or_admin(self, comp: Competition) -> None: sender = self._sender() if comp.created_by == sender: return if sender in self.admins and self.admins[sender]: return raise gl.vm.UserError( f"{ERROR_EXPECTED} Only the competition creator or an admin may do this" ) def _log(self, event: str, data: dict) -> None: # Keep audit entries compact; storage is not free. try: payload = json.dumps( {"e": event, "d": data}, sort_keys=True, separators=(",", ":") ) except (TypeError, ValueError): payload = json.dumps({"e": event, "d": "unserializable"}) if len(payload) > 800: payload = payload[:800] self.audit_log.append(payload) def _get_weights(self) -> dict: try: weights = json.loads(self.weights_json) except (ValueError, TypeError): weights = dict(DEFAULT_WEIGHTS_BP) # Guarantee every criterion has a weight even after config mistakes. for criterion in CRITERIA: if criterion not in weights: weights[criterion] = DEFAULT_WEIGHTS_BP[criterion] return weights def _get_competition(self, competition_id: str) -> Competition: if competition_id not in self.competitions: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition not found: {competition_id}" ) return self.competitions[competition_id] def _get_submission(self, submission_id: str) -> Submission: if submission_id not in self.submissions: raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission not found: {submission_id}" ) return self.submissions[submission_id] def _comp_submission_ids(self, competition_id: str) -> list: if competition_id not in self.competition_submissions_json: return [] raw = self.competition_submissions_json[competition_id] try: ids = json.loads(raw) return ids if isinstance(ids, list) else [] except (ValueError, TypeError): return [] def _append_comp_submission(self, competition_id: str, submission_id: str) -> None: ids = self._comp_submission_ids(competition_id) ids.append(submission_id) self.competition_submissions_json[competition_id] = json.dumps( ids, separators=(",", ":") ) @staticmethod def _timestamp_from_iso(value: str, field: str) -> int: """Parse a user-supplied ISO-8601 timestamp deterministically.""" try: normalized = value.replace("Z", "+00:00") parsed = datetime.fromisoformat(normalized) if parsed.tzinfo is None: raise ValueError("timezone required") return int(parsed.astimezone(timezone.utc).timestamp()) except (TypeError, ValueError, OverflowError): raise gl.vm.UserError(f"{ERROR_EXPECTED} {field} must be ISO-8601 with timezone") @staticmethod def _now_epoch() -> int: # GenVM pins datetime.now() to the transaction datetime, so this is # consensus-safe and cannot be forged by a direct caller. return int(datetime.now(timezone.utc).timestamp()) def _has_open_dispute(self, competition_id: str) -> bool: for did in self.dispute_ids: dispute = self.disputes[did] if dispute.competition_id == competition_id and dispute.status == DISPUTE_STATUS_OPEN: return True return False @staticmethod def _normalize_url(url: str) -> str: u = url.strip().lower() u = re.sub(r"^https?://", "", u) u = re.sub(r"^www\.", "", u) u = u.split("?")[0].split("#")[0].rstrip("/") return u @staticmethod def _validate_url(url: str, field: str) -> None: if not url or len(url) > MAX_URL_LEN: raise gl.vm.UserError( f"{ERROR_EXPECTED} {field} must be 1..{MAX_URL_LEN} chars" ) if not re.match(r"^https?://[^\s]+\.[^\s]+", url): raise gl.vm.UserError( f"{ERROR_EXPECTED} {field} must be a valid http(s) URL" ) def _touch_user(self, address: str) -> UserStats: if address not in self.user_stats: self.user_stats[address] = UserStats( address=address, submissions=u256(0), wins=u256(0), total_score_accum=u256(0), rewards_atto=u256(0), disqualifications=u256(0), last_submission_at="", ) return self.user_stats[address] # ------------------------------------------------------------------ # Nondeterministic helpers — only callable inside nondet closures. # ------------------------------------------------------------------ @staticmethod def _web_get(url: str): """Fetch a URL inside a nondet block, classifying failures.""" try: response = gl.nondet.web.get(url) except Exception as exc: # network layer failure — transient raise gl.vm.UserError(f"{ERROR_TRANSIENT} web fetch failed: {exc}") status = getattr(response, "status", 0) if 400 <= status < 500: raise gl.vm.UserError(f"{ERROR_EXTERNAL} HTTP {status} for resource") if status >= 500: raise gl.vm.UserError(f"{ERROR_TRANSIENT} HTTP {status} upstream error") return response @staticmethod def _web_text(url: str) -> str: """Render a page as text inside a nondet block, tolerant to SDK API surface differences between runner versions.""" render = getattr(getattr(gl.nondet, "web", None), "render", None) if render is not None: try: return str(render(url, mode="text")) except TypeError: pass get_webpage = getattr(gl, "get_webpage", None) if get_webpage is not None: return str(get_webpage(url, mode="text")) raise gl.vm.UserError(f"{ERROR_TRANSIENT} no web render capability") @staticmethod def _exec_prompt_json(prompt: str, images: list = None) -> dict: """Run an LLM prompt and coerce the response into a dict. `images`, if given, is a list of raw image bytes — genuine visual input, not just a URL string in the prompt text. NOTE: the runtime's own type stubs are inconsistent between overloads (one shows `image=` singular for JSON mode) but the actual implementation only reads the `images` (plural, list) kwarg regardless of response_format — confirmed by testing; passing `image=` singular is silently a no-op.""" try: if images: result = gl.nondet.exec_prompt( prompt, response_format="json", images=images ) else: result = gl.nondet.exec_prompt(prompt, response_format="json") except Exception as exc: raise gl.vm.UserError(f"{ERROR_LLM} exec_prompt failed: {exc}") if isinstance(result, dict): return result if isinstance(result, str): parsed = MemeOlympics._parse_json_text(result) if parsed is not None: return parsed raise gl.vm.UserError( f"{ERROR_LLM} LLM returned non-JSON response: {type(result)}" ) @staticmethod def _parse_json_text(text: str): """Clean typical LLM JSON wrapping: prose, code fences, trailing commas.""" try: first = text.find("{") last = text.rfind("}") if first == -1 or last == -1 or last <= first: return None snippet = text[first : last + 1] snippet = re.sub(r",(\s*[}\]])", r"\1", snippet) parsed = json.loads(snippet) return parsed if isinstance(parsed, dict) else None except (ValueError, TypeError): return None @staticmethod def _coerce_int(value, lo: int, hi: int, field: str) -> int: """Aggressively coerce an LLM-provided number into a clamped int.""" if value is None: raise gl.vm.UserError(f"{ERROR_LLM} missing numeric field '{field}'") try: number = int(round(float(str(value).strip()))) except (ValueError, TypeError): raise gl.vm.UserError( f"{ERROR_LLM} non-numeric field '{field}': {value}" ) return max(lo, min(hi, number)) @staticmethod def _pick(analysis: dict, key: str, aliases: tuple): """Fetch a key from an LLM dict with common alias fallbacks.""" if key in analysis: return analysis[key] for alias in aliases: if alias in analysis: return analysis[alias] return None @staticmethod def _handle_leader_error(leaders_res, leader_fn) -> bool: """Canonical validator-side comparison when the leader errored. Deterministic errors must match exactly; transient errors agree if both sides are transient; LLM/unknown errors force disagreement so consensus rotates the leader instead of persisting garbage. """ leader_msg = getattr(leaders_res, "message", "") or "" try: leader_fn() return False # leader failed, validator succeeded -> disagree except gl.vm.UserError as exc: validator_msg = getattr(exc, "message", None) or str(exc) if validator_msg.startswith(ERROR_EXPECTED) or validator_msg.startswith( ERROR_EXTERNAL ): return validator_msg == leader_msg if validator_msg.startswith(ERROR_TRANSIENT) and leader_msg.startswith( ERROR_TRANSIENT ): return True return False except Exception: return False # ================================================================== # Admin & configuration (deterministic writes) # ================================================================== @gl.public.write def claim_initial_owner(self) -> None: """One-time bootstrap for a known deployment quirk: some deploy paths present a zero sender to the constructor (and/or drop constructor args), leaving self.owner as the zero address. If that happened, the first caller of this regular (non-constructor) method becomes owner/admin — sender_address is reliable on ordinary calls. A no-op safety door: rejected once a real owner already exists.""" zero = Address("0x0000000000000000000000000000000000000000") if self.owner != zero: raise gl.vm.UserError(f"{ERROR_EXPECTED} Owner already set") self.owner = gl.message.sender_address self.admins[self._addr_str(self.owner)] = True self._log("owner_claimed", {"owner": self._addr_str(self.owner)}) @gl.public.write def transfer_ownership(self, new_owner: str) -> None: self._require_owner() new_addr = Address(new_owner) self.owner = new_addr self.admins[self._addr_str(new_addr)] = True self._log("ownership_transferred", {"to": new_owner}) @gl.public.write def add_admin(self, admin_address: str) -> None: self._require_owner() self.admins[Address(admin_address).as_hex] = True self._log("admin_added", {"admin": admin_address}) @gl.public.write def remove_admin(self, admin_address: str) -> None: self._require_owner() key = Address(admin_address).as_hex if key == self._addr_str(self.owner): raise gl.vm.UserError(f"{ERROR_EXPECTED} Cannot remove owner from admins") if key in self.admins: self.admins[key] = False self._log("admin_removed", {"admin": admin_address}) @gl.public.write def set_paused(self, paused: bool) -> None: self._require_admin() self.paused = paused self._log("paused_set", {"paused": paused}) @gl.public.write def set_weights(self, weights_json: str) -> None: """Update criterion weights (basis points). Must cover all criteria and sum to 10000.""" self._require_admin() try: weights = json.loads(weights_json) except (ValueError, TypeError): raise gl.vm.UserError(f"{ERROR_EXPECTED} weights_json is not valid JSON") if not isinstance(weights, dict): raise gl.vm.UserError(f"{ERROR_EXPECTED} weights_json must be an object") total = 0 for criterion in CRITERIA: if criterion not in weights: raise gl.vm.UserError( f"{ERROR_EXPECTED} missing weight for '{criterion}'" ) try: w = int(weights[criterion]) except (ValueError, TypeError): raise gl.vm.UserError( f"{ERROR_EXPECTED} weight for '{criterion}' must be an integer" ) if w < 0 or w > 10000: raise gl.vm.UserError( f"{ERROR_EXPECTED} weight for '{criterion}' out of range" ) total += w if total != 10000: raise gl.vm.UserError( f"{ERROR_EXPECTED} weights must sum to 10000, got {total}" ) self.weights_json = json.dumps( {c: int(weights[c]) for c in CRITERIA}, sort_keys=True ) self._log("weights_updated", {}) @gl.public.write def set_competition_defaults(self, winner_count: int, max_per_user: int) -> None: """Default winner count / per-user submission cap for NEW competitions. Prize pools are no longer a config default — they are a declared USDC amount set per-competition via create_competition's `prize_pool_usdc` or fund_competition, backed by a real deposit on the Base Sepolia escrow contract.""" self._require_admin() if winner_count < 1 or winner_count > 25: raise gl.vm.UserError(f"{ERROR_EXPECTED} winner_count must be 1..25") if max_per_user < 1 or max_per_user > 20: raise gl.vm.UserError(f"{ERROR_EXPECTED} max_per_user must be 1..20") self.default_winner_count = u256(winner_count) self.max_submissions_per_user = u256(max_per_user) self._log( "defaults_updated", {"winners": winner_count, "mpu": max_per_user} ) # ================================================================== # Competition lifecycle (deterministic writes) # ================================================================== @gl.public.write def create_competition( self, competition_id: str, title: str, theme: str, starts_at: str, ends_at: str, prize_pool_usdc: int = 0, ) -> None: """Create a competition. OPEN TO ALL, unconditionally — anyone can host an arena, with no lifetime cap and no admin approval ever required. IDs are caller-supplied (e.g. 'week-2026-28') so the off-chain scheduler and the chain agree on naming without extra reads. Spam is bounded off-chain instead (a time-windowed rate limit that resets automatically, unlike a hard on-chain cap that would need an admin bypass) — see backend rate limiting on this route. NOT PAYABLE: this contract never escrows real value. `prize_pool_usdc` (USDC base units, 6 decimals) is a declared amount for display only — the real prize money is deposited separately by the host on the Base Sepolia escrow contract under this same competition_id (0 is valid — a prestige-only arena with no monetary reward). Anyone can bump the declared amount later via fund_competition.""" self._require_not_paused() if prize_pool_usdc < 0: raise gl.vm.UserError(f"{ERROR_EXPECTED} prize_pool_usdc must be >= 0") funded = u256(prize_pool_usdc) sender = self._sender() # Track creation count for stats only — no cap, no admin bypass needed. creator_key = f"created_by:{sender}" created_count = int(self.user_comp_submission_count[creator_key]) if ( creator_key in self.user_comp_submission_count ) else 0 self.user_comp_submission_count[creator_key] = u256(created_count + 1) if not re.match(r"^[a-z0-9][a-z0-9\-]{2,63}$", competition_id): raise gl.vm.UserError( f"{ERROR_EXPECTED} competition_id must be 3..64 chars of [a-z0-9-]" ) if competition_id in self.competitions: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition already exists: {competition_id}" ) if not title or len(title) > MAX_TITLE_LEN: raise gl.vm.UserError(f"{ERROR_EXPECTED} title must be 1..{MAX_TITLE_LEN}") if len(theme) > MAX_CAPTION_LEN: raise gl.vm.UserError(f"{ERROR_EXPECTED} theme too long") start_ts = self._timestamp_from_iso(starts_at, "starts_at") end_ts = self._timestamp_from_iso(ends_at, "ends_at") if end_ts <= start_ts: raise gl.vm.UserError(f"{ERROR_EXPECTED} ends_at must be after starts_at") self.competitions[competition_id] = Competition( id=competition_id, title=title, theme=theme, status=COMP_STATUS_CREATED, starts_at=starts_at, ends_at=ends_at, prize_pool_atto=u256(0), winner_count=self.default_winner_count, submission_count=u256(0), created_by=self._sender(), created_at=starts_at, finalized_at="", winners_json="[]", prize_pool_usdc=funded, relay_tx_hash="", challenge_ends_at=u256(0), ) self.competition_ids.append(competition_id) self.total_competitions = u256(int(self.total_competitions) + 1) self._log( "competition_created", {"id": competition_id, "declared_prize_usdc": int(funded)}, ) @gl.public.write def fund_competition(self, competition_id: str, added_usdc: int) -> None: """Bump a competition's declared USDC prize pool. Open to anyone — the host, sponsors, or the community — as long as the arena hasn't finalized or been cancelled yet. NOT PAYABLE: still no real value moves here — deposit the matching USDC on the Base Sepolia escrow contract under this competition_id.""" self._require_not_paused() comp = self._get_competition(competition_id) if comp.status in (COMP_STATUS_FINALIZED, COMP_STATUS_CANCELLED): raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition '{comp.status}' can no longer be funded" ) if added_usdc <= 0: raise gl.vm.UserError(f"{ERROR_EXPECTED} added_usdc must be > 0") comp.prize_pool_usdc = u256(int(comp.prize_pool_usdc) + added_usdc) self._log( "competition_funded", {"id": competition_id, "added_usdc": added_usdc}, ) @gl.public.write def open_competition(self, competition_id: str) -> None: """Open a competition for submissions and make it the active one. Allowed for the competition's creator or any admin.""" self._require_not_paused() comp = self._get_competition(competition_id) self._require_creator_or_admin(comp) if comp.status != COMP_STATUS_CREATED: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition is '{comp.status}', expected 'created'" ) # Concurrent arenas are allowed: opening one never closes another. # The active pointer is only a "featured" hint for UIs. comp.status = COMP_STATUS_OPEN self.active_competition_id = competition_id self._log("competition_opened", {"id": competition_id}) @gl.public.write def close_submissions(self, competition_id: str) -> None: """Move an open competition into the judging phase. Allowed for the competition's creator or any admin.""" comp = self._get_competition(competition_id) self._require_creator_or_admin(comp) if comp.status != COMP_STATUS_OPEN: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition is '{comp.status}', expected 'open'" ) if self._now_epoch() < self._timestamp_from_iso(comp.ends_at, "ends_at"): raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition deadline has not passed" ) comp.status = COMP_STATUS_JUDGING if self.active_competition_id == competition_id: self.active_competition_id = "" self._log("competition_judging", {"id": competition_id}) @gl.public.write def cancel_competition(self, competition_id: str, reason: str) -> None: self._require_admin() comp = self._get_competition(competition_id) if comp.status == COMP_STATUS_FINALIZED: raise gl.vm.UserError( f"{ERROR_EXPECTED} Finalized competitions cannot be cancelled" ) comp.status = COMP_STATUS_CANCELLED if self.active_competition_id == competition_id: self.active_competition_id = "" self._log("competition_cancelled", {"id": competition_id, "r": reason[:200]}) # ================================================================== # Submissions (deterministic write) # ================================================================== @gl.public.write def submit_meme( self, competition_id: str, submission_id: str, title: str, caption: str, image_url: str, context_url: str, tags_json: str, submitted_at: str, ) -> None: """Register a meme submission for the given open competition. The heavy AI judging happens later in `evaluate_submission`; registration itself is deterministic and cheap so submitting never risks consensus failure.""" self._require_not_paused() comp = self._get_competition(competition_id) if comp.status != COMP_STATUS_OPEN: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition '{competition_id}' is not open" ) if int(comp.submission_count) >= MAX_SUBMISSIONS_PER_COMPETITION: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition submission cap reached" ) if not re.match(r"^[a-zA-Z0-9][a-zA-Z0-9\-_]{2,63}$", submission_id): raise gl.vm.UserError( f"{ERROR_EXPECTED} submission_id must be 3..64 url-safe chars" ) if submission_id in self.submissions: raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission already exists: {submission_id}" ) if not title or len(title) > MAX_TITLE_LEN: raise gl.vm.UserError( f"{ERROR_EXPECTED} title must be 1..{MAX_TITLE_LEN} chars" ) if len(caption) > MAX_CAPTION_LEN: raise gl.vm.UserError(f"{ERROR_EXPECTED} caption too long") self._validate_url(image_url, "image_url") if context_url: self._validate_url(context_url, "context_url") # Tag validation. try: tags = json.loads(tags_json) if tags_json else [] except (ValueError, TypeError): raise gl.vm.UserError(f"{ERROR_EXPECTED} tags_json is not valid JSON") if not isinstance(tags, list) or len(tags) > MAX_TAGS: raise gl.vm.UserError(f"{ERROR_EXPECTED} tags must be a list of <= {MAX_TAGS}") clean_tags = [] for tag in tags: tag_str = str(tag).strip()[:MAX_TAG_LEN] if tag_str: clean_tags.append(tag_str) sender = self._sender() # Per-user submission cap for this competition. quota_key = f"{competition_id}:{sender}" used = int(self.user_comp_submission_count[quota_key]) if ( quota_key in self.user_comp_submission_count ) else 0 if used >= int(self.max_submissions_per_user): raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission limit reached for this competition" ) # Exact-duplicate image URL detection (first line of anti-recycling; # the AI evaluation performs the deeper originality analysis). norm = self._normalize_url(image_url) if norm in self.image_url_seen: raise gl.vm.UserError( f"{ERROR_EXPECTED} This image URL was already submitted " f"(submission {self.image_url_seen[norm]})" ) self.submissions[submission_id] = Submission( id=submission_id, competition_id=competition_id, author=sender, title=title, caption=caption, image_url=image_url, context_url=context_url, tags_json=json.dumps(clean_tags, separators=(",", ":")), status=SUB_STATUS_PENDING, submitted_at=submitted_at, total_score=u256(0), criteria_json="{}", plagiarism_verdict="", plagiarism_confidence=u256(0), image_accessible=False, evaluation_summary="", evaluated_at="", eligible=True, eligibility_reason="", ) self.submission_ids.append(submission_id) self._append_comp_submission(competition_id, submission_id) self.image_url_seen[norm] = submission_id self.user_comp_submission_count[quota_key] = u256(used + 1) comp.submission_count = u256(int(comp.submission_count) + 1) self.total_submissions = u256(int(self.total_submissions) + 1) stats = self._touch_user(sender) stats.submissions = u256(int(stats.submissions) + 1) stats.last_submission_at = submitted_at self._log("meme_submitted", {"id": submission_id, "c": competition_id}) # ================================================================== # AI EVALUATION — the consensus core. # ================================================================== def _build_evaluation_prompt( self, sub: Submission, theme: str, page_note: str, has_image: bool = False ) -> str: weights = self._get_weights() criteria_lines = "\n".join( f'- "{c}" (weight {weights[c]} bp): score 0-10' for c in CRITERIA ) vision_note = ( "The actual meme image is attached below — you can genuinely SEE " "it. Base your judgment on what is literally depicted (subject, " "template/format, composition, text overlay, image quality), not " "just the metadata." if has_image else "The image could not be attached this time — judge from the " "metadata below only, and treat that as a limitation, not a " "confident visual assessment." ) return f"""You are an expert judge for "Meme Olympics", a weekly crypto meme competition. Judge the following meme submission on reasoning-based criteria, not popularity. Be consistent, fair and moderately generous: reserve scores below 3 for clearly poor or off-topic entries, and scores above 8 for genuinely exceptional ones. {vision_note} COMPETITION THEME: {theme or "open theme — current crypto culture"} SUBMISSION METADATA: - Title: {sub.title} - Caption: {sub.caption} - Tags: {sub.tags_json} - Context URL provided: {"yes" if sub.context_url else "no"} - Image accessibility check: {page_note} TASKS: 0. Eligibility gate — answer BEFORE scoring. Set "eligible" to false ONLY if the submission fails to even be a real meme entry, e.g.: the image is blank/corrupted/unrelated stock content with no meme composition (text overlay, template, juxtaposition, etc.), the title/caption is empty or pure gibberish/spam, or the content is obviously unrelated to memes entirely (an ad, a random product photo, a test upload). Do NOT set it false just because the meme is low-quality, unfunny, or off-theme — that is reflected in the criteria scores below, not the gate. Give a short "eligibility_reason" either way (e.g. "valid meme entry" when eligible). 1. Score each criterion as an integer 0-10 (still fill this in even if not eligible — use your honest best-effort scores; the contract will zero the final score when the gate fails): {criteria_lines} 2. Plagiarism / recycling assessment. Look at the actual image content: is this a well-known meme template (e.g. Drake, Distracted Boyfriend, Woman Yelling at Cat) reused with fresh text/context — that's NORMAL and fine — or does it look like a direct screenshot/repost of existing content with someone else's caption/watermark still on it, or a near-identical duplicate of something widely circulated? Combine what you SEE with the title/caption and your knowledge of widely circulated memes on Reddit, X/Twitter, Telegram, Discord, 4chan and meme archives, then classify: - "original": a novel creation or a substantially transformative remix - "suspicious": a known template/format with limited transformation (NORMAL for memes — using a known template with fresh text/context is fine and should usually remain scoreable) - "copied": very likely a direct repost/screenshot of existing content with no meaningful transformation (e.g. visible original watermark, screenshot chrome, or pixel-identical to a widely known image) Give a confidence 0-100 for your classification. Only choose "copied" when the visual evidence is strong; when in doubt between suspicious and copied, choose "suspicious". 3. Write a 1-3 sentence judging summary. Return ONLY a JSON object exactly like: {{ "eligible": true, "eligibility_reason": "...", "criteria": {{ {", ".join(f'"{c}": 0' for c in CRITERIA)} }}, "plagiarism_verdict": "original|suspicious|copied", "plagiarism_confidence": 0, "summary": "..." }}""" def _run_evaluation(self, sub: Submission, theme: str) -> dict: """Leader/validator shared task: probe the image URL, then run the LLM judging prompt — WITH the actual image bytes as genuine visual input, not just a URL string — and normalize the result to stable fields.""" # 1) Contract-side web access: verify the meme image actually exists, # and capture its bytes for real visual analysis (confirmed working # via live testing: the LLM correctly identifies actual pixel # content — colors, subjects, scenes — not just filename/URL text). image_accessible = False image_bytes = None page_note = "unknown" try: response = self._web_get(sub.image_url) image_accessible = True image_bytes = getattr(response, "body", None) content_type = "" headers = getattr(response, "headers", None) if isinstance(headers, dict): for k, v in headers.items(): if str(k).lower() == "content-type": content_type = str(v) break page_note = f"reachable (content-type: {content_type or 'n/a'})" except gl.vm.UserError as exc: msg = getattr(exc, "message", None) or str(exc) if msg.startswith(ERROR_TRANSIENT): raise # let transient failures classify as transient page_note = "NOT reachable (HTTP client error)" # 2) LLM judging — visual when we have image bytes, text-only # fallback otherwise (e.g. fetch failed but wasn't disqualifying). images = [image_bytes] if isinstance(image_bytes, (bytes, bytearray)) else None analysis = self._exec_prompt_json( self._build_evaluation_prompt(sub, theme, page_note, bool(images)), images=images, ) # 3) Defensive normalization to stable decision fields. eligible_raw = self._pick(analysis, "eligible", ("is_eligible", "valid")) eligible = bool(eligible_raw) if eligible_raw is not None else True eligibility_reason = str( self._pick(analysis, "eligibility_reason", ("reason", "ineligible_reason")) or ("" if eligible else "not a valid meme entry") )[:300] raw_criteria = self._pick(analysis, "criteria", ("scores", "ratings")) if not isinstance(raw_criteria, dict): raise gl.vm.UserError(f"{ERROR_LLM} missing criteria object") criteria_scores = {} for criterion in CRITERIA: value = raw_criteria.get(criterion) if value is None: # tolerate alias variants like "cryptoNativeUnderstanding" for k, v in raw_criteria.items(): if re.sub(r"[^a-z]", "", str(k).lower()) == re.sub( r"[^a-z]", "", criterion ): value = v break criteria_scores[criterion] = self._coerce_int( value if value is not None else 5, 0, 10, criterion ) verdict = str( self._pick(analysis, "plagiarism_verdict", ("verdict", "plagiarism")) or PLAGIARISM_SUSPICIOUS ).strip().lower() if verdict not in ( PLAGIARISM_ORIGINAL, PLAGIARISM_SUSPICIOUS, PLAGIARISM_COPIED, ): verdict = PLAGIARISM_SUSPICIOUS confidence = self._coerce_int( self._pick(analysis, "plagiarism_confidence", ("confidence",)) or 50, 0, 100, "plagiarism_confidence", ) weights = self._get_weights() total_bp = sum(criteria_scores[c] * weights[c] for c in CRITERIA) # Hard gate: an ineligible submission scores 0 regardless of what the # per-axis numbers say — axes never rescue a disqualified entry. total_score = 0 if not eligible else max(0, min(100, total_bp // 1000)) summary = str( self._pick(analysis, "summary", ("analysis", "reasoning")) or "" )[:500] return { "eligible": eligible, "eligibility_reason": eligibility_reason, "criteria": criteria_scores, "total_score": int(total_score), "plagiarism_verdict": verdict, "plagiarism_confidence": int(confidence), "image_accessible": bool(image_accessible), "summary": summary, } @gl.public.write def evaluate_submission(self, submission_id: str) -> None: """Run the AI judging pipeline for one submission under consensus. The leader performs image probing + LLM scoring; every validator INDEPENDENTLY re-runs the same task and compares the substantive decision fields with explicit tolerances. Validators never accept a leader result on format alone.""" self._require_not_paused() sub = self._get_submission(submission_id) comp = self._get_competition(sub.competition_id) if comp.status != COMP_STATUS_JUDGING: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition is '{comp.status}'; cannot evaluate" ) if sub.status not in (SUB_STATUS_PENDING,): raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission is '{sub.status}'; already processed" ) theme = comp.theme # Snapshot plain values for the nondet closures (avoid touching # storage objects inside nondet blocks). sub_snapshot = Submission( id=sub.id, competition_id=sub.competition_id, author=sub.author, title=sub.title, caption=sub.caption, image_url=sub.image_url, context_url=sub.context_url, tags_json=sub.tags_json, status=sub.status, submitted_at=sub.submitted_at, total_score=u256(0), criteria_json="{}", plagiarism_verdict="", plagiarism_confidence=u256(0), image_accessible=False, evaluation_summary="", evaluated_at="", eligible=True, eligibility_reason="", ) contract = self def leader_fn() -> dict: return contract._run_evaluation(sub_snapshot, theme) def validator_fn(leaders_res: gl.vm.Result) -> bool: if not isinstance(leaders_res, gl.vm.Return): return MemeOlympics._handle_leader_error(leaders_res, leader_fn) leader = leaders_res.calldata if not isinstance(leader, dict): return False # Independent re-derivation — the validator does the task itself. try: mine = leader_fn() except gl.vm.UserError as exc: msg = getattr(exc, "message", None) or str(exc) # Leader succeeded but validator hit transient trouble: # accept the leader rather than rotating for network noise. return msg.startswith(ERROR_TRANSIENT) except Exception: return False # --- Gate 0: eligibility is a hard gate. Both sides must agree # on whether this is even a scoreable meme entry at all — this # is checked BEFORE plagiarism/score/criteria, and no amount of # axis agreement below can rescue a mismatch here. leader_eligible = bool(leader.get("eligible", True)) if leader_eligible != mine["eligible"]: return False # --- Gate 1: plagiarism "copied" is a hard gate. Both sides # must agree on whether the meme is disqualified. "original" vs # "suspicious" disagreement is tolerated (does not disqualify). leader_verdict = str(leader.get("plagiarism_verdict", "")) leader_conf = leader.get("plagiarism_confidence", 0) try: leader_conf = int(leader_conf) except (ValueError, TypeError): return False leader_dq = ( leader_verdict == PLAGIARISM_COPIED and leader_conf >= 70 ) my_dq = ( mine["plagiarism_verdict"] == PLAGIARISM_COPIED and mine["plagiarism_confidence"] >= 70 ) if leader_dq != my_dq: # Soft escape hatch: near-threshold confidence differences # should not rotate the leader forever. If verdicts are both # copied-ish (copied/suspicious) treat as agreement. both_copied_ish = leader_verdict in ( PLAGIARISM_COPIED, PLAGIARISM_SUSPICIOUS, ) and mine["plagiarism_verdict"] in ( PLAGIARISM_COPIED, PLAGIARISM_SUSPICIOUS, ) if not both_copied_ish: return False # --- Gate 1b: when either side flagged anything but a clean # "original", require the two confidence numbers to actually be # close — otherwise a leader could claim "copied, 71% confident" # against a validator's "copied, 5% confident" and still pass # gate 1 above purely because both crossed the same threshold. both_original = ( leader_verdict == PLAGIARISM_ORIGINAL and mine["plagiarism_verdict"] == PLAGIARISM_ORIGINAL ) if not both_original: if abs(leader_conf - mine["plagiarism_confidence"]) > CONFIDENCE_TOLERANCE: return False # --- Gate 2: total score tolerance. try: leader_total = int(leader.get("total_score", -1)) except (ValueError, TypeError): return False if leader_total < 0 or leader_total > 100: return False if abs(leader_total - mine["total_score"]) > SCORE_TOLERANCE: return False # --- Gate 3: per-criterion sanity. Every leader score must be a # valid 0..10 int and within tolerance of the validator's own # score for at least CRITERIA_AGREEMENT_MIN of 9 criteria. leader_criteria = leader.get("criteria") if not isinstance(leader_criteria, dict): return False close_enough = 0 for criterion in CRITERIA: try: lv = int(leader_criteria.get(criterion, -1)) except (ValueError, TypeError): return False if lv < 0 or lv > 10: return False if abs(lv - mine["criteria"][criterion]) <= CRITERION_TOLERANCE: close_enough += 1 if close_enough < CRITERIA_AGREEMENT_MIN: return False # --- Gate 4: image accessibility must match (it is a near- # deterministic fact; disagreement means someone's fetch lied). if bool(leader.get("image_accessible")) != mine["image_accessible"]: # tolerate flaky hosting: only fail when validator says the # image exists but leader claimed it does not if mine["image_accessible"] and not leader.get("image_accessible"): return False return True result = gl.vm.run_nondet_unsafe(leader_fn, validator_fn) # ------------------------------------------------------------------ # Deterministic state update from the agreed result. # ------------------------------------------------------------------ sub.criteria_json = json.dumps(result["criteria"], sort_keys=True) sub.total_score = u256(int(result["total_score"])) sub.plagiarism_verdict = str(result["plagiarism_verdict"]) sub.plagiarism_confidence = u256(int(result["plagiarism_confidence"])) sub.image_accessible = bool(result["image_accessible"]) sub.evaluation_summary = str(result["summary"])[:500] sub.evaluated_at = sub.submitted_at # chain has no clock; see docs sub.eligible = bool(result["eligible"]) sub.eligibility_reason = str(result["eligibility_reason"])[:300] disqualified = ( result["plagiarism_verdict"] == PLAGIARISM_COPIED and int(result["plagiarism_confidence"]) >= 70 ) or not result["image_accessible"] if not sub.eligible: # Gate 0 failure — never scoreable, distinct from a disqualified # (plagiarized/unreachable) entry that at least passed the # eligibility bar. Kept separate so hosts/UI can tell "wasn't a # real meme" apart from "was a meme but broke the rules". sub.status = SUB_STATUS_REJECTED stats = self._touch_user(sub.author) stats.disqualifications = u256(int(stats.disqualifications) + 1) elif disqualified: sub.status = SUB_STATUS_DISQUALIFIED stats = self._touch_user(sub.author) stats.disqualifications = u256(int(stats.disqualifications) + 1) else: sub.status = SUB_STATUS_EVALUATED stats = self._touch_user(sub.author) stats.total_score_accum = u256( int(stats.total_score_accum) + int(result["total_score"]) ) self.total_evaluations = u256(int(self.total_evaluations) + 1) self._log( "submission_evaluated", { "id": submission_id, "s": int(result["total_score"]), "v": result["plagiarism_verdict"], "dq": disqualified, }, ) # ================================================================== # Finalization & rewards (fully deterministic) # ================================================================== @gl.public.write def record_unjudgeable_attempt(self, submission_id: str, evidence_tx_hash: str) -> None: """Commit one distinct failed evaluation transaction as evidence. This does not change submission status and is independently auditable.""" self._require_admin() sub = self._get_submission(submission_id) comp = self._get_competition(sub.competition_id) if comp.status != COMP_STATUS_JUDGING: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition is '{comp.status}', expected 'judging'" ) if sub.status != SUB_STATUS_PENDING: raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission is '{sub.status}', expected 'pending'" ) evidence = str(evidence_tx_hash).strip().lower() if len(evidence) != 66 or not evidence.startswith("0x"): raise gl.vm.UserError(f"{ERROR_EXPECTED} Invalid evaluation transaction hash") try: int(evidence[2:], 16) except ValueError: raise gl.vm.UserError(f"{ERROR_EXPECTED} Invalid evaluation transaction hash") hashes_json = ( self.unjudgeable_attempt_hashes_json[submission_id] if submission_id in self.unjudgeable_attempt_hashes_json else "[]" ) hashes = json.loads(hashes_json or "[]") if evidence in hashes: raise gl.vm.UserError(f"{ERROR_EXPECTED} Evidence transaction already recorded") if len(hashes) >= 3: raise gl.vm.UserError(f"{ERROR_EXPECTED} Required evidence already complete") hashes.append(evidence) self.unjudgeable_attempt_hashes_json[submission_id] = json.dumps(hashes) self.unjudgeable_attempt_count[submission_id] = u256(len(hashes)) self._log( "unjudgeable_attempt_recorded", {"id": submission_id, "attempt": len(hashes), "tx": evidence}, ) @gl.public.write def disqualify_unjudgeable_submission(self, submission_id: str, reason: str) -> None: """Owner-only terminal recovery for an entry the AI runtime cannot decode after bounded retries. This is deliberately narrower than a general moderation endpoint: only a pending entry in a judging arena can transition, and the reason is persisted in the audit log/state.""" self._require_owner() sub = self._get_submission(submission_id) comp = self._get_competition(sub.competition_id) if comp.status != COMP_STATUS_JUDGING: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition is '{comp.status}', expected 'judging'" ) if sub.status != SUB_STATUS_PENDING: raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission is '{sub.status}', expected 'pending'" ) evidence_count = ( int(self.unjudgeable_attempt_count[submission_id]) if submission_id in self.unjudgeable_attempt_count else 0 ) if evidence_count < 3: raise gl.vm.UserError( f"{ERROR_EXPECTED} Three failed evaluation commitments are required" ) clean_reason = str(reason).strip()[:300] if not clean_reason: raise gl.vm.UserError(f"{ERROR_EXPECTED} Disqualification reason is required") sub.status = SUB_STATUS_DISQUALIFIED sub.criteria_json = "{}" sub.total_score = u256(0) sub.image_accessible = False sub.evaluation_summary = clean_reason sub.evaluated_at = sub.submitted_at sub.eligible = False sub.eligibility_reason = clean_reason stats = self._touch_user(sub.author) stats.disqualifications = u256(int(stats.disqualifications) + 1) self.total_evaluations = u256(int(self.total_evaluations) + 1) self._log( "submission_terminally_disqualified", {"id": submission_id, "reason": clean_reason, "evidence_count": evidence_count}, ) @gl.public.write def finalize_competition(self, competition_id: str, finalized_at: str) -> None: """Deterministically rank evaluated submissions, select winners and compute each one's USDC reward share. Purely a judging-side record — no value moves here; the backend relayer reads this result and calls the Base Sepolia escrow contract to make the payout claimable there (see `mark_prizes_relayed`). No LLM/web calls — cannot fail consensus. Allowed for the competition's creator or any admin, so hosts can end their own arena without waiting on an admin.""" comp = self._get_competition(competition_id) self._require_creator_or_admin(comp) if comp.status != COMP_STATUS_JUDGING: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition is '{comp.status}', expected 'judging'" ) sub_ids = self._comp_submission_ids(competition_id) # Never rank a convenient subset. Every registered entry must have a # terminal judging result before winner selection is allowed. for sid in sub_ids: if sid not in self.submissions: raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission record missing: {sid}" ) status = self.submissions[sid].status if status not in ( SUB_STATUS_EVALUATED, SUB_STATUS_REJECTED, SUB_STATUS_DISQUALIFIED, ): raise gl.vm.UserError( f"{ERROR_EXPECTED} Submission '{sid}' is '{status}'; judging incomplete" ) if self._has_open_dispute(competition_id): raise gl.vm.UserError( f"{ERROR_EXPECTED} Cannot finalize while disputes are unresolved" ) winners = self._assign_winners(comp) comp.status = COMP_STATUS_FINALIZED # Do not trust the caller's wall-clock string for settlement timing. # Keep the parameter for ABI compatibility, but persist consensus time. now = self._now_epoch() comp.finalized_at = datetime.fromtimestamp(now, timezone.utc).isoformat() comp.challenge_ends_at = u256(now + CHALLENGE_PERIOD_SECONDS) if self.active_competition_id == competition_id: self.active_competition_id = "" self._log( "competition_finalized", {"id": competition_id, "winners": len(winners), "challenge_ends_at": int(comp.challenge_ends_at)}, ) @staticmethod def _reward_shares(winner_count: int) -> list: """Descending integer shares in basis points summing to 10000.""" if winner_count <= 0: return [] if winner_count == 1: return [10000] if winner_count == 2: return [6000, 4000] if winner_count == 3: return [5000, 3000, 2000] # Generalized: weight n-i for rank i, normalized to 10000 bp. weights = [winner_count - i for i in range(winner_count)] total = sum(weights) shares = [w * 10000 // total for w in weights] shares[0] += 10000 - sum(shares) # dust to rank 1 return shares def _assign_winners(self, comp: Competition) -> list: """Rank all currently eligible entries and create the provisional winners list. Call only after any old winner allocation has been removed; this is also used after an upheld challenge to promote the next eligible entry before Base settlement.""" ranked = [] for sid in self._comp_submission_ids(comp.id): if sid in self.submissions: sub = self.submissions[sid] if sub.status == SUB_STATUS_EVALUATED: ranked.append((-int(sub.total_score), sid)) ranked.sort() winner_count = min(int(comp.winner_count), len(ranked)) shares = self._reward_shares(winner_count) winners = [] for index in range(winner_count): sub = self.submissions[ranked[index][1]] reward = int(comp.prize_pool_usdc) * shares[index] // 10000 sub.status = SUB_STATUS_WINNER stats = self._touch_user(sub.author) stats.wins = u256(int(stats.wins) + 1) stats.rewards_atto = u256(int(stats.rewards_atto) + reward) balance = int(self.reward_balances_atto[sub.author]) if sub.author in self.reward_balances_atto else 0 self.reward_balances_atto[sub.author] = u256(balance + reward) self.total_rewards_distributed_atto = u256(int(self.total_rewards_distributed_atto) + reward) winners.append({ "submission_id": sub.id, "author": sub.author, "rank": index + 1, "score": int(sub.total_score), "reward_usdc": str(reward), }) comp.winners_json = json.dumps(winners, separators=(",", ":")) return winners # ================================================================== # Payout relay — NO value moves on this chain. This contract only # records that a competition's winners were pushed to the Base Sepolia # USDC escrow, so the backend relayer job is idempotent and the frontend # can show payout status without re-deriving it off-chain. # ================================================================== @gl.public.write def mark_prizes_relayed(self, competition_id: str, relay_tx_hash: str) -> None: """Admin/relayer-only marker: call this once the winners list for `competition_id` has been submitted to the Base Sepolia escrow contract (MemeOlympicsEscrow.setWinners) and that transaction has confirmed. Moves no value — winners actually claim their USDC directly from the escrow contract on Base Sepolia.""" self._require_admin() comp = self._get_competition(competition_id) if comp.status != COMP_STATUS_FINALIZED: raise gl.vm.UserError( f"{ERROR_EXPECTED} Competition is '{comp.status}', expected 'finalized'" ) if self._now_epoch() < int(comp.challenge_ends_at): raise gl.vm.UserError( f"{ERROR_EXPECTED} Challenge period is still open" ) if self._has_open_dispute(competition_id): raise gl.vm.UserError( f"{ERROR_EXPECTED} Cannot relay prizes while disputes are unresolved" ) if comp.relay_tx_hash: raise gl.vm.UserError( f"{ERROR_EXPECTED} Already marked relayed: {comp.relay_tx_hash}" ) if not relay_tx_hash: raise gl.vm.UserError(f"{ERROR_EXPECTED} relay_tx_hash required") comp.relay_tx_hash = relay_tx_hash self.total_rewards_claimed_atto = u256( int(self.total_rewards_claimed_atto) + int(comp.prize_pool_usdc) ) self._log( "prizes_relayed", {"id": competition_id, "tx": relay_tx_hash} ) # ================================================================== # Disputes — evidence-based, contract-side web verification # ================================================================== @gl.public.write def open_dispute( self, dispute_id: str, submission_id: str, reason: str, evidence_url: str, opened_at: str, ) -> None: """Open a plagiarism/misjudgment dispute against a submission. The disputer must supply a public evidence URL (e.g. a Reddit or X post showing the meme existed earlier). Resolution fetches that evidence ON-CHAIN — user-submitted text alone never decides.""" self._require_not_paused() sub = self._get_submission(submission_id) comp = self._get_competition(sub.competition_id) if ( comp.status == COMP_STATUS_FINALIZED and self._now_epoch() >= int(comp.challenge_ends_at) ): raise gl.vm.UserError( f"{ERROR_EXPECTED} Challenge period has ended" ) if sub.status not in (SUB_STATUS_EVALUATED, SUB_STATUS_WINNER): raise gl.vm.UserError( f"{ERROR_EXPECTED} Only evaluated or winning submissions can be disputed" ) if not re.match(r"^[a-zA-Z0-9][a-zA-Z0-9\-_]{2,63}$", dispute_id): raise gl.vm.UserError(f"{ERROR_EXPECTED} invalid dispute_id") if dispute_id in self.disputes: raise gl.vm.UserError(f"{ERROR_EXPECTED} Dispute already exists") if not reason or len(reason) > MAX_DISPUTE_REASON_LEN: raise gl.vm.UserError( f"{ERROR_EXPECTED} reason must be 1..{MAX_DISPUTE_REASON_LEN} chars" ) self._validate_url(evidence_url, "evidence_url") count_key = submission_id existing = int(self.submission_dispute_count[count_key]) if ( count_key in self.submission_dispute_count ) else 0 if existing >= 3: raise gl.vm.UserError( f"{ERROR_EXPECTED} Dispute limit reached for this submission" ) self.disputes[dispute_id] = Dispute( id=dispute_id, submission_id=submission_id, competition_id=sub.competition_id, opened_by=self._sender(), reason=reason, evidence_url=evidence_url, status=DISPUTE_STATUS_OPEN, verdict_summary="", opened_at=opened_at, resolved_at="", ) self.dispute_ids.append(dispute_id) self.submission_dispute_count[count_key] = u256(existing + 1) self.total_disputes = u256(int(self.total_disputes) + 1) self._log("dispute_opened", {"id": dispute_id, "sub": submission_id}) def _run_dispute_analysis( self, dispute: Dispute, sub: Submission ) -> dict: """Shared leader/validator task for dispute resolution: fetch the evidence page on-chain, then reason over it.""" evidence_excerpt = "" evidence_fetched = False try: text = self._web_text(dispute.evidence_url) evidence_excerpt = text[:4000] evidence_fetched = True except gl.vm.UserError as exc: msg = getattr(exc, "message", None) or str(exc) if msg.startswith(ERROR_TRANSIENT): raise evidence_excerpt = "(evidence URL could not be fetched: client error)" prompt = f"""You are resolving a dispute in a crypto meme competition. DISPUTED SUBMISSION: - Title: {sub.title} - Caption: {sub.caption} - Image URL: {sub.image_url} - Current plagiarism verdict: {sub.plagiarism_verdict} \ (confidence {int(sub.plagiarism_confidence)}) - Score: {int(sub.total_score)}/100 DISPUTE: - Claim by disputer: {dispute.reason} - Evidence URL: {dispute.evidence_url} - Evidence page fetched on-chain: {"yes" if evidence_fetched else "NO — could not fetch"} EVIDENCE PAGE CONTENT (first 4000 chars): {evidence_excerpt} TASK: Decide whether the dispute should be UPHELD (the evidence genuinely demonstrates the claim, e.g. the meme was published elsewhere earlier or the judgment was clearly wrong) or REJECTED. IMPORTANT RULES: - If the evidence page could not be fetched, you MUST reject: claims cannot be resolved from the disputer's text alone. - Only uphold when the evidence content concretely supports the claim. - When genuinely uncertain, reject (submissions keep the benefit of doubt). Return ONLY JSON: {{"uphold": true/false, "confidence": 0-100, "summary": "1-2 sentences"}}""" analysis = self._exec_prompt_json(prompt) uphold_raw = self._pick(analysis, "uphold", ("upheld", "decision", "verdict")) if isinstance(uphold_raw, str): uphold = uphold_raw.strip().lower() in ("true", "uphold", "upheld", "yes") else: uphold = bool(uphold_raw) if not evidence_fetched: uphold = False # hard rule mirrored in code, not just the prompt confidence = self._coerce_int( self._pick(analysis, "confidence", ("certainty",)) or 50, 0, 100, "confidence" ) summary = str(self._pick(analysis, "summary", ("reasoning",)) or "")[:400] return { "uphold": uphold, "confidence": confidence, "summary": summary, "evidence_fetched": evidence_fetched, } @gl.public.write def resolve_dispute(self, dispute_id: str, resolved_at: str) -> None: """Resolve a dispute under consensus. Validators independently fetch the same evidence and re-derive the verdict; agreement is on the uphold/reject decision, not prose.""" self._require_not_paused() if dispute_id not in self.disputes: raise gl.vm.UserError(f"{ERROR_EXPECTED} Dispute not found: {dispute_id}") dispute = self.disputes[dispute_id] if dispute.status != DISPUTE_STATUS_OPEN: raise gl.vm.UserError(f"{ERROR_EXPECTED} Dispute already resolved") sub = self._get_submission(dispute.submission_id) comp = self._get_competition(sub.competition_id) dispute_snapshot = Dispute( id=dispute.id, submission_id=dispute.submission_id, competition_id=dispute.competition_id, opened_by=dispute.opened_by, reason=dispute.reason, evidence_url=dispute.evidence_url, status=dispute.status, verdict_summary="", opened_at=dispute.opened_at, resolved_at="", ) sub_snapshot = Submission( id=sub.id, competition_id=sub.competition_id, author=sub.author, title=sub.title, caption=sub.caption, image_url=sub.image_url, context_url=sub.context_url, tags_json=sub.tags_json, status=sub.status, submitted_at=sub.submitted_at, total_score=sub.total_score, criteria_json=sub.criteria_json, plagiarism_verdict=sub.plagiarism_verdict, plagiarism_confidence=sub.plagiarism_confidence, image_accessible=sub.image_accessible, evaluation_summary="", evaluated_at="", eligible=sub.eligible, eligibility_reason="", ) contract = self def leader_fn() -> dict: return contract._run_dispute_analysis(dispute_snapshot, sub_snapshot) def validator_fn(leaders_res: gl.vm.Result) -> bool: if not isinstance(leaders_res, gl.vm.Return): return MemeOlympics._handle_leader_error(leaders_res, leader_fn) leader = leaders_res.calldata if not isinstance(leader, dict): return False try: mine = leader_fn() except gl.vm.UserError as exc: msg = getattr(exc, "message", None) or str(exc) return msg.startswith(ERROR_TRANSIENT) except Exception: return False # Decision gate: uphold/reject must match… if bool(leader.get("uphold")) != mine["uphold"]: # …unless both are low-confidence borderline calls, where we # defer to the leader to avoid rotation loops on 50/50 cases. try: leader_conf = int(leader.get("confidence", 0)) except (ValueError, TypeError): return False if leader_conf >= 65 or mine["confidence"] >= 65: return False # Evidence-fetched must not be contradicted in the strict # direction (leader claims fetched, validator could not is # tolerable network variance; the reverse suggests a lying leader # only when leader also upheld). if ( mine["evidence_fetched"] and not bool(leader.get("evidence_fetched")) and bool(leader.get("uphold")) ): return False return True result = gl.vm.run_nondet_unsafe(leader_fn, validator_fn) # Deterministic settlement. if result["uphold"]: dispute.status = DISPUTE_STATUS_UPHELD previous_status = sub.status sub.status = SUB_STATUS_DISQUALIFIED sub.plagiarism_verdict = PLAGIARISM_COPIED stats = self._touch_user(sub.author) stats.disqualifications = u256(int(stats.disqualifications) + 1) # A provisional winner can be disqualified only while settlement # is still locked. Re-rank the entire eligible field so the next # valid entry is promoted and the full prize allocation remains # fair before anything can be claimed on Base. if previous_status == SUB_STATUS_WINNER: self._rebalance_provisional_winners(comp) else: dispute.status = DISPUTE_STATUS_REJECTED dispute.verdict_summary = str(result["summary"])[:400] dispute.resolved_at = resolved_at self._log( "dispute_resolved", {"id": dispute_id, "u": result["uphold"]}, ) def _rebalance_provisional_winners(self, comp: Competition) -> None: """Undo the previous provisional allocation and assign a fresh, deterministic winners list. `mark_prizes_relayed` is impossible while a dispute is open, so no Base allocation can exist at this point.""" try: winners = json.loads(comp.winners_json) except (ValueError, TypeError): winners = [] for w in winners: sid = str(w.get("submission_id", "")) if sid not in self.submissions: continue old_winner = self.submissions[sid] reward = int(w.get("reward_usdc", w.get("reward_atto", "0"))) balance = int(self.reward_balances_atto[old_winner.author]) if old_winner.author in self.reward_balances_atto else 0 self.reward_balances_atto[old_winner.author] = u256(max(0, balance - reward)) stats = self._touch_user(old_winner.author) stats.rewards_atto = u256(max(0, int(stats.rewards_atto) - reward)) stats.wins = u256(max(0, int(stats.wins) - 1)) self.total_rewards_distributed_atto = u256(max(0, int(self.total_rewards_distributed_atto) - reward)) if old_winner.status == SUB_STATUS_WINNER: old_winner.status = SUB_STATUS_EVALUATED self._assign_winners(comp) # ================================================================== # Views — read-only, no gas surprises # ================================================================== @gl.public.view def get_contract_info(self) -> dict: return { "name": "MemeOlympics", "version": "1.0.0", "owner": self._addr_str(self.owner), "paused": self.paused, "active_competition_id": self.active_competition_id, "total_competitions": int(self.total_competitions), "total_submissions": int(self.total_submissions), "total_evaluations": int(self.total_evaluations), "total_disputes": int(self.total_disputes), "total_prizes_declared_usdc": str( int(self.total_rewards_distributed_atto) ), "total_prizes_relayed_usdc": str(int(self.total_rewards_claimed_atto)), "payment_chain": "base-sepolia", "payment_asset": "USDC", "criteria": list(CRITERIA), } @gl.public.view def get_config(self) -> dict: return { "weights_bp": self._get_weights(), "default_winner_count": int(self.default_winner_count), "max_submissions_per_user": int(self.max_submissions_per_user), "max_submissions_per_competition": MAX_SUBMISSIONS_PER_COMPETITION, "score_tolerance": SCORE_TOLERANCE, } @gl.public.view def is_admin(self, address: str) -> bool: key = Address(address).as_hex return bool(key in self.admins and self.admins[key]) @gl.public.view def get_competition(self, competition_id: str) -> dict: comp = self._get_competition(competition_id) return { "id": comp.id, "title": comp.title, "theme": comp.theme, "status": comp.status, "starts_at": comp.starts_at, "ends_at": comp.ends_at, "prize_pool_usdc": str(int(comp.prize_pool_usdc)), "winner_count": int(comp.winner_count), "submission_count": int(comp.submission_count), "created_by": comp.created_by, "finalized_at": comp.finalized_at, "winners": json.loads(comp.winners_json or "[]"), "relay_tx_hash": comp.relay_tx_hash, "challenge_ends_at": int(comp.challenge_ends_at), "challenge_period_seconds": CHALLENGE_PERIOD_SECONDS, } @gl.public.view def can_relay_prizes(self, competition_id: str) -> bool: """True only when the contract itself would accept the relay marker.""" comp = self._get_competition(competition_id) return ( comp.status == COMP_STATUS_FINALIZED and not comp.relay_tx_hash and self._now_epoch() >= int(comp.challenge_ends_at) and not self._has_open_dispute(competition_id) ) @gl.public.view def get_active_competition(self) -> dict: if not self.active_competition_id: return {"active": False} result = self.get_competition(self.active_competition_id) result["active"] = True return result @gl.public.view def list_competitions(self, offset: int, limit: int) -> list: """Paginated newest-first competition summaries.""" limit = max(1, min(50, limit)) offset = max(0, offset) total = len(self.competition_ids) out = [] index = total - 1 - offset while index >= 0 and len(out) < limit: cid = self.competition_ids[index] comp = self.competitions[cid] out.append( { "id": comp.id, "title": comp.title, "status": comp.status, "starts_at": comp.starts_at, "ends_at": comp.ends_at, "submission_count": int(comp.submission_count), } ) index -= 1 return out @gl.public.view def get_submission(self, submission_id: str) -> dict: sub = self._get_submission(submission_id) return { "id": sub.id, "competition_id": sub.competition_id, "author": sub.author, "title": sub.title, "caption": sub.caption, "image_url": sub.image_url, "context_url": sub.context_url, "tags": json.loads(sub.tags_json or "[]"), "status": sub.status, "submitted_at": sub.submitted_at, "total_score": int(sub.total_score), "criteria": json.loads(sub.criteria_json or "{}"), "plagiarism_verdict": sub.plagiarism_verdict, "plagiarism_confidence": int(sub.plagiarism_confidence), "image_accessible": sub.image_accessible, "evaluation_summary": sub.evaluation_summary, "eligible": sub.eligible, "eligibility_reason": sub.eligibility_reason, "unjudgeable_attempt_count": ( int(self.unjudgeable_attempt_count[submission_id]) if submission_id in self.unjudgeable_attempt_count else 0 ), "unjudgeable_attempt_tx_hashes": json.loads( ( self.unjudgeable_attempt_hashes_json[submission_id] if submission_id in self.unjudgeable_attempt_hashes_json else "[]" ) or "[]" ), } @gl.public.view def list_competition_submissions(self, competition_id: str) -> list: self._get_competition(competition_id) out = [] for sid in self._comp_submission_ids(competition_id): if sid not in self.submissions: continue sub = self.submissions[sid] out.append( { "id": sub.id, "author": sub.author, "title": sub.title, "image_url": sub.image_url, "status": sub.status, "total_score": int(sub.total_score), "plagiarism_verdict": sub.plagiarism_verdict, } ) return out @gl.public.view def get_leaderboard(self, competition_id: str) -> list: """Score-ranked evaluated/winner submissions for one competition.""" self._get_competition(competition_id) entries = [] for sid in self._comp_submission_ids(competition_id): if sid not in self.submissions: continue sub = self.submissions[sid] if sub.status in (SUB_STATUS_EVALUATED, SUB_STATUS_WINNER): entries.append((-int(sub.total_score), sid)) entries.sort() out = [] rank = 1 for negative_score, sid in entries: sub = self.submissions[sid] out.append( { "rank": rank, "submission_id": sid, "author": sub.author, "title": sub.title, "image_url": sub.image_url, "score": -negative_score, "status": sub.status, } ) rank += 1 return out @gl.public.view def get_user_stats(self, address: str) -> dict: key = Address(address).as_hex if key not in self.user_stats: return { "address": key, "submissions": 0, "wins": 0, "total_score_accum": 0, "rewards_usdc": "0", "disqualifications": 0, } stats = self.user_stats[key] return { "address": stats.address, "submissions": int(stats.submissions), "wins": int(stats.wins), "total_score_accum": int(stats.total_score_accum), "rewards_usdc": str(int(stats.rewards_atto)), "disqualifications": int(stats.disqualifications), } @gl.public.view def get_declared_reward(self, address: str) -> str: """Total USDC (base units) this address has been declared as owed across all won competitions. Informational only — the amount is actually claimable on Base Sepolia only once the relevant competition has been `mark_prizes_relayed` onto the escrow contract; check that competition's `relay_tx_hash` first.""" key = Address(address).as_hex if key not in self.reward_balances_atto: return "0" return str(int(self.reward_balances_atto[key])) @gl.public.view def get_dispute(self, dispute_id: str) -> dict: if dispute_id not in self.disputes: raise gl.vm.UserError(f"{ERROR_EXPECTED} Dispute not found: {dispute_id}") d = self.disputes[dispute_id] return { "id": d.id, "submission_id": d.submission_id, "competition_id": d.competition_id, "opened_by": d.opened_by, "reason": d.reason, "evidence_url": d.evidence_url, "status": d.status, "verdict_summary": d.verdict_summary, "opened_at": d.opened_at, "resolved_at": d.resolved_at, } @gl.public.view def list_disputes(self, offset: int, limit: int) -> list: limit = max(1, min(50, limit)) offset = max(0, offset) total = len(self.dispute_ids) out = [] index = total - 1 - offset while index >= 0 and len(out) < limit: did = self.dispute_ids[index] d = self.disputes[did] out.append( { "id": d.id, "submission_id": d.submission_id, "status": d.status, "opened_at": d.opened_at, } ) index -= 1 return out @gl.public.view def get_audit_log_tail(self, count: int) -> list: """Return up to the last `count` audit events (capped at 100).""" count = max(1, min(100, count)) total = len(self.audit_log) start = max(0, total - count) out = [] for index in range(start, total): try: out.append(json.loads(self.audit_log[index])) except (ValueError, TypeError): out.append({"e": "unparseable"}) return out