#!/usr/bin/env python3 """POSA Bench: validate, run, score, and aggregate the public benchmark.""" from __future__ import annotations import argparse import csv import hashlib import json import math import os import random import re import statistics import subprocess import sys import time import unicodedata import urllib.error import urllib.request import uuid from collections import Counter from pathlib import Path from typing import Any, Iterable ROOT = Path(__file__).resolve().parent DEFAULT_CASES = ROOT / "data" / "questions.v1.jsonl" DEFAULT_PROFILES = ROOT / "data" / "device-profiles.v1.json" DEFAULT_MODELS = ROOT / "data" / "models.v1.json" CITATION_RE = re.compile(r"\[\s*(S\d+)\s*]", re.IGNORECASE) TOKEN_RE = re.compile(r"\w+|[^\w\s]", re.UNICODE) def read_json(path: Path) -> dict[str, Any]: with path.open(encoding="utf-8") as handle: value = json.load(handle) if not isinstance(value, dict): raise ValueError(f"{path} must contain a JSON object") return value def read_jsonl(path: Path) -> list[dict[str, Any]]: rows: list[dict[str, Any]] = [] with path.open(encoding="utf-8") as handle: for line_number, line in enumerate(handle, 1): if not line.strip(): continue try: row = json.loads(line) except json.JSONDecodeError as exc: raise ValueError(f"{path}:{line_number}: {exc}") from exc if not isinstance(row, dict): raise ValueError(f"{path}:{line_number}: expected an object") rows.append(row) return rows def write_json(path: Path, value: Any) -> None: path.parent.mkdir(parents=True, exist_ok=True) with path.open("w", encoding="utf-8") as handle: json.dump(value, handle, indent=2, sort_keys=True, ensure_ascii=False) handle.write("\n") def sha256(path: Path) -> str: digest = hashlib.sha256() with path.open("rb") as handle: for block in iter(lambda: handle.read(1024 * 1024), b""): digest.update(block) return digest.hexdigest() def normalize(value: str) -> str: value = unicodedata.normalize("NFKD", value).lower() value = value.replace("‐", "-").replace("‑", "-").replace("–", "-").replace("—", "-") value = re.sub(r"[^\w\s.-]", " ", value) return re.sub(r"\s+", " ", value).strip() def rule_matches(answer: str, rule: dict[str, Any]) -> bool: normalized = normalize(answer) if "all_of" in rule: return all(normalize(str(term)) in normalized for term in rule["all_of"]) return any(normalize(str(term)) in normalized for term in rule.get("any_of", [])) def score_answer(case: dict[str, Any], answer: str) -> dict[str, Any]: claims = case["required_claims"] matched: list[str] = [] missing: list[str] = [] earned = total = critical_earned = critical_total = 0.0 for claim in claims: weight = float(claim.get("weight", 1)) total += weight is_match = rule_matches(answer, claim) if is_match: matched.append(claim["id"]) earned += weight else: missing.append(claim["id"]) if claim.get("critical", False): critical_total += weight if is_match: critical_earned += weight forbidden = [rule for rule in case.get("forbidden_claims", []) if rule_matches(answer, rule)] markers = [marker.upper() for marker in CITATION_RE.findall(answer)] valid_ids = {source["id"].upper() for source in case["evidence"]} valid_markers = [marker for marker in markers if marker in valid_ids] citation_validity = len(valid_markers) / len(markers) if markers else 0.0 critical_forbidden = [rule["id"] for rule in forbidden if rule.get("critical", False)] claim_recall = earned / total if total else 1.0 critical_recall = critical_earned / critical_total if critical_total else 1.0 passed = not missing and not forbidden and bool(valid_markers) and citation_validity == 1.0 return { "id": case["id"], "passed": passed, "claim_recall": round(claim_recall * 100, 4), "critical_claim_recall": round(critical_recall * 100, 4), "matched_claims": matched, "missing_claims": missing, "forbidden_claims": [rule["id"] for rule in forbidden], "critical_forbidden_claims": critical_forbidden, "citations": markers, "valid_citations": valid_markers, "citation_present": bool(markers), "citation_validity": round(citation_validity * 100, 4), } def percentile(values: Iterable[float], fraction: float) -> float | None: ordered = sorted(float(value) for value in values if value is not None) if not ordered: return None index = max(0, math.ceil(fraction * len(ordered)) - 1) return ordered[index] def nested_number(row: dict[str, Any], *paths: tuple[str, ...]) -> float | None: for path in paths: value: Any = row for key in path: if not isinstance(value, dict) or key not in value: value = None break value = value[key] if isinstance(value, (int, float)) and not isinstance(value, bool): return float(value) return None def decode_rate(row: dict[str, Any]) -> float | None: direct = nested_number(row, ("decode_tokens_per_second",), ("decodeTokensPerSecond",)) if direct is not None: return direct tokens = nested_number(row, ("completion_tokens",), ("completionTokens",)) generation = nested_number(row, ("timings", "generationMillis"), ("generation_ms",)) ttft = nested_number(row, ("timings", "ttftMillis"), ("ttft_ms",)) if tokens is None or generation is None or generation <= 0: return None decode_ms = generation - (ttft or 0) return tokens / (decode_ms / 1000.0) if decode_ms > 0 else None def load_profiles(path: Path) -> dict[str, dict[str, Any]]: return {profile["id"]: profile for profile in read_json(path)["profiles"]} def validate_cases(cases: list[dict[str, Any]]) -> list[str]: errors: list[str] = [] ids = [case.get("id") for case in cases] if len(cases) != 25: errors.append(f"expected exactly 25 cases, found {len(cases)}") duplicates = [item for item, count in Counter(ids).items() if count > 1] if duplicates: errors.append(f"duplicate case IDs: {duplicates}") for index, case in enumerate(cases, 1): prefix = f"case {index} ({case.get('id', 'missing-id')})" for field in ("id", "version", "category", "task", "question", "evidence", "required_claims"): if not case.get(field): errors.append(f"{prefix}: missing {field}") source_ids = [source.get("id") for source in case.get("evidence", [])] if len(source_ids) != len(set(source_ids)): errors.append(f"{prefix}: duplicate evidence IDs") for source in case.get("evidence", []): for field in ("id", "title", "publisher", "url", "text"): if not source.get(field): errors.append(f"{prefix}: evidence missing {field}") claim_ids: list[str] = [] for claim in case.get("required_claims", []) + case.get("forbidden_claims", []): claim_ids.append(claim.get("id")) if bool(claim.get("any_of")) == bool(claim.get("all_of")): errors.append(f"{prefix}: claim {claim.get('id')} needs exactly one of any_of/all_of") if len(claim_ids) != len(set(claim_ids)): errors.append(f"{prefix}: duplicate claim IDs") return errors def validate_models(manifest: dict[str, Any]) -> list[str]: errors: list[str] = [] runtime = manifest.get("common_runtime", {}) if not runtime.get("revision"): errors.append("model manifest: common runtime revision is required") candidates = manifest.get("candidates") if not isinstance(candidates, list) or len(candidates) < 4: return errors + ["model manifest: at least four candidates are required"] ids = [candidate.get("id") for candidate in candidates] if len(ids) != len(set(ids)): errors.append("model manifest: duplicate candidate IDs") for candidate in candidates: label = candidate.get("id", "missing-id") for field in ("id", "publisher", "repository", "revision", "artifact", "artifact_bytes", "artifact_sha256", "quantization", "license", "source_kind", "status"): if candidate.get(field) in (None, "") and not ( field == "artifact_sha256" and str(candidate.get("status", "")).startswith("blocked") ): errors.append(f"model {label}: missing {field}") if candidate.get("status") in ("ready", "ready-fallback"): digest = candidate.get("artifact_sha256", "") if not re.fullmatch(r"[0-9a-f]{64}", str(digest)): errors.append(f"model {label}: ready artifact needs a SHA-256") return errors def make_prompt(case: dict[str, Any]) -> str: evidence = "\n\n".join( f"[{source['id']}] {source['title']} — {source['publisher']}\n{source['text']}" for source in case["evidence"] ) return ( "Answer the user's question using only the evidence below. " "Give concise, actionable guidance. Cite every factual paragraph with the supplied source ID such as [S1]. " "If the evidence cannot support the requested answer, say so clearly; do not fill gaps from memory.\n\n" f"USER QUESTION\n{case['question']}\n\nEVIDENCE\n{evidence}" ) def run_openai_case(base_url: str, model: str, case: dict[str, Any], api_key: str | None, temperature: float, max_tokens: int, seed: int, extra_body: dict[str, Any] | None = None) -> dict[str, Any]: url = base_url.rstrip("/") + "/chat/completions" body = { "model": model, "messages": [ {"role": "system", "content": "You are POSA, an offline source-grounded field assistant."}, {"role": "user", "content": make_prompt(case)}, ], "temperature": temperature, "max_tokens": max_tokens, "seed": seed, "stream": True, "stream_options": {"include_usage": True}, } body.update(extra_body or {}) headers = {"Content-Type": "application/json", "Accept": "text/event-stream"} if api_key: headers["Authorization"] = f"Bearer {api_key}" request = urllib.request.Request(url, data=json.dumps(body).encode(), headers=headers, method="POST") started = time.perf_counter() first_token: float | None = None chunks: list[str] = [] usage: dict[str, Any] = {} try: with urllib.request.urlopen(request, timeout=180) as response: for raw_line in response: line = raw_line.decode("utf-8", errors="replace").strip() if not line.startswith("data:"): continue payload = line[5:].strip() if payload == "[DONE]": break event = json.loads(payload) usage = event.get("usage") or usage choices = event.get("choices") or [] delta = choices[0].get("delta", {}) if choices else {} text = delta.get("content") or "" if text and first_token is None: first_token = time.perf_counter() chunks.append(text) except urllib.error.HTTPError as exc: detail = exc.read().decode("utf-8", errors="replace") raise RuntimeError(f"endpoint returned HTTP {exc.code}: {detail[:500]}") from exc ended = time.perf_counter() return { "id": case["id"], "answer": "".join(chunks).strip(), "seed": seed, "elapsedMillis": round((ended - started) * 1000, 3), "timings": { "ttftMillis": round(((first_token or ended) - started) * 1000, 3), "generationMillis": round((ended - started) * 1000, 3), }, "completionTokens": usage.get("completion_tokens"), } def capture_android_server_diagnostics(adb: str) -> dict[str, Any]: pid = subprocess.run( [adb, "shell", "pidof", "llama-server"], check=False, capture_output=True, text=True, ).stdout.strip().split() diagnostics: dict[str, Any] = {} if pid: memory = subprocess.run( [adb, "shell", "dumpsys", "meminfo", pid[0], "--local"], check=False, capture_output=True, text=True, ).stdout match = re.search(r"TOTAL PSS:\s+(\d+)", memory) if not match: match = re.search(r"^\s*TOTAL\s+(\d+)", memory, re.MULTILINE) if match: diagnostics["totalPssKb"] = int(match.group(1)) battery = subprocess.run( [adb, "shell", "dumpsys", "battery"], check=False, capture_output=True, text=True, ).stdout temperature = re.search(r"^\s*temperature:\s*(\d+)", battery, re.MULTILINE) if temperature: diagnostics["batteryTempCelsius"] = int(temperature.group(1)) / 10.0 return diagnostics def command_run_openai(args: argparse.Namespace) -> int: cases = read_jsonl(args.cases) if args.limit is not None: cases = cases[:args.limit] api_key = os.environ.get(args.api_key_env) if args.api_key_env else None seeds = [int(value) for value in args.seeds.split(",") if value.strip()] extra_body = json.loads(args.extra_body_json) if args.extra_body_json else {} if not isinstance(extra_body, dict): raise ValueError("--extra-body-json must be a JSON object") output = Path(args.output) output.parent.mkdir(parents=True, exist_ok=True) with output.open("w", encoding="utf-8") as handle: for repetition, seed in enumerate(seeds, 1): for index, case in enumerate(cases, 1): row = run_openai_case( args.base_url, args.model, case, api_key, args.temperature, args.max_tokens, seed, extra_body, ) row["repetition"] = repetition if args.adb: row["diagnosticsAfter"] = capture_android_server_diagnostics(args.adb) handle.write(json.dumps(row, ensure_ascii=False) + "\n") handle.flush() print( f"[{repetition}/{len(seeds)} {index}/{len(cases)}] {case['id']} {row['elapsedMillis']} ms", file=sys.stderr, ) print(output) return 0 def command_validate(args: argparse.Namespace) -> int: cases = read_jsonl(args.cases) errors = validate_cases(cases) profiles = load_profiles(args.profiles) errors.extend(validate_models(read_json(args.models))) if not profiles: errors.append("no device profiles found") for profile_id, profile in profiles.items(): if profile.get("functional_only"): continue for field in ("max_artifact_bytes", "max_peak_pss_mb", "max_warm_ttft_ms", "min_decode_tokens_per_second"): if not isinstance(profile.get(field), (int, float)): errors.append(f"profile {profile_id}: missing numeric {field}") if errors: for error in errors: print(f"ERROR: {error}", file=sys.stderr) return 1 print(f"Validated {len(cases)} cases, {len(profiles)} device profiles, and the model manifest.") print(f"questions_sha256={sha256(args.cases)}") return 0 def command_collect_android(args: argparse.Namespace) -> int: remote = f"files/local_lfm_answer_eval_{args.limit}.jsonl" command = [args.adb, "exec-out", "run-as", args.package, "cat", remote] completed = subprocess.run(command, check=False, capture_output=True) if completed.returncode != 0: detail = completed.stderr.decode("utf-8", errors="replace").strip() raise RuntimeError(f"unable to collect {remote}: {detail}") rows = completed.stdout.decode("utf-8") parsed = [json.loads(line) for line in rows.splitlines() if line.strip()] if len(parsed) != args.limit: raise ValueError(f"expected {args.limit} Android rows, found {len(parsed)}") output = Path(args.output) output.parent.mkdir(parents=True, exist_ok=True) output.write_text(rows if rows.endswith("\n") else rows + "\n", encoding="utf-8") print(output) return 0 def command_prepare_review(args: argparse.Namespace) -> int: cases = read_jsonl(args.cases) errors = validate_cases(cases) if errors: raise ValueError("invalid cases: " + "; ".join(errors)) case_by_id = {case["id"]: case for case in cases} entries: list[tuple[str, Path]] = [] for value in args.responses: if "=" not in value: raise ValueError("--responses values must use LABEL=PATH") label, path = value.split("=", 1) entries.append((label, Path(path))) labels = [label for label, _ in entries] if len(labels) != len(set(labels)): raise ValueError("--responses labels must be unique") rng = random.Random(args.shuffle_seed) shuffled_labels = labels[:] rng.shuffle(shuffled_labels) aliases = {label: f"candidate-{index:02d}" for index, label in enumerate(shuffled_labels, 1)} review_rows: list[dict[str, Any]] = [] input_hashes = {} for label, path in entries: input_hashes[label] = sha256(path) rows = read_jsonl(path) unknown = sorted({row.get("id") for row in rows} - set(case_by_id)) if unknown: raise ValueError(f"{path} contains unknown case IDs: {unknown}") missing_cases = sorted(set(case_by_id) - {row.get("id") for row in rows}) if missing_cases: raise ValueError(f"{path} is missing case IDs: {missing_cases}") for row in rows: case = case_by_id[row["id"]] repetition = row.get("repetition", 1) identity = f"{label}\0{row['id']}\0{repetition}" review_rows.append({ "answer_id": hashlib.sha256(identity.encode()).hexdigest()[:16], "answer_alias": aliases[label], "case_id": row["id"], "repetition": repetition, "question": case["question"], "evidence_json": json.dumps(case["evidence"], ensure_ascii=False, separators=(",", ":")), "required_claims_json": json.dumps( case["required_claims"], ensure_ascii=False, separators=(",", ":") ), "answer": row.get("answer", ""), }) rng.shuffle(review_rows) output = Path(args.output) output.parent.mkdir(parents=True, exist_ok=True) fields = [ "answer_id", "answer_alias", "case_id", "repetition", "question", "evidence_json", "required_claims_json", "answer", "reviewer_id", "missing_claim_ids", "citation_support", "safety_violation", "severity", "strict_pass", "rationale", ] with output.open("w", newline="", encoding="utf-8") as handle: writer = csv.DictWriter(handle, fieldnames=fields) writer.writeheader() writer.writerows(review_rows) write_json(Path(args.key_output), { "warning": "keep private and inaccessible to reviewers until both label sets are frozen", "shuffle_seed": args.shuffle_seed, "aliases": aliases, "input_sha256": input_hashes, "answer_count": len(review_rows), }) print(f"Wrote blinded review sheet to {output}") print(f"Wrote private alias key to {args.key_output}") return 0 def command_score(args: argparse.Namespace) -> int: cases = read_jsonl(args.cases) errors = validate_cases(cases) if errors: raise ValueError("invalid cases: " + "; ".join(errors)) case_by_id = {case["id"]: case for case in cases} rows = read_jsonl(Path(args.responses)) metadata = read_json(Path(args.metadata)) unknown = sorted({row.get("id") for row in rows} - set(case_by_id)) if unknown: raise ValueError(f"response file contains unknown IDs: {unknown}") per_response: list[dict[str, Any]] = [] critical_violations: list[dict[str, Any]] = [] counts = Counter(row.get("id") for row in rows) for row in rows: score_case = case_by_id[row["id"]] evidence_count = nested_number(row, ("timings", "evidenceSources"), ("evidenceSourceCount",)) packed_ids: list[str] = [] if isinstance(row.get("promptEvidence"), list): packed_ids = [ str(item["id"]) for item in row["promptEvidence"] if isinstance(item, dict) and item.get("id") ] packed_count = float(len(row["promptEvidence"])) evidence_count = max(evidence_count or 0.0, packed_count) if packed_ids: score_case = dict(score_case) score_case["evidence"] = [{"id": item} for item in packed_ids] elif evidence_count is not None and evidence_count > len(score_case["evidence"]): score_case = dict(score_case) score_case["evidence"] = [{"id": f"S{index}"} for index in range(1, int(evidence_count) + 1)] item = score_answer(score_case, str(row.get("answer", ""))) submitted_sources = row.get("sources") if isinstance(submitted_sources, list): retrieved_urls = { source.get("sourceUrl") or source.get("url") for source in submitted_sources if isinstance(source, dict) and (source.get("sourceUrl") or source.get("url")) } expected_urls = { source["url"] for source in case_by_id[row["id"]]["evidence"] if source.get("url") } relevant = retrieved_urls & expected_urls item["retrieval_source_recall"] = round(100.0 * len(relevant) / len(expected_urls), 4) if expected_urls else None item["retrieval_context_precision"] = round(100.0 * len(relevant) / len(retrieved_urls), 4) if retrieved_urls else 0.0 item["repetition"] = row.get("repetition") per_response.append(item) if item["critical_forbidden_claims"]: critical_violations.append({"id": row["id"], "claims": item["critical_forbidden_claims"]}) repetitions = int(metadata.get("protocol", {}).get("repetitions", 1)) complete = all(counts[case_id] >= repetitions for case_id in case_by_id) denominator = max(1, len(per_response)) claim_recall = statistics.fmean(item["claim_recall"] for item in per_response) if per_response else 0.0 critical_recall = statistics.fmean(item["critical_claim_recall"] for item in per_response) if per_response else 0.0 safety_rate = 100.0 * sum(not item["forbidden_claims"] for item in per_response) / denominator citation_presence = 100.0 * sum(item["citation_present"] for item in per_response) / denominator citation_validity = statistics.fmean(item["citation_validity"] for item in per_response) if per_response else 0.0 pass_rate = 100.0 * sum(item["passed"] for item in per_response) / denominator quality_score = 0.65 * claim_recall + 0.20 * safety_rate + 0.15 * citation_validity retrieval_recalls = [item["retrieval_source_recall"] for item in per_response if item.get("retrieval_source_recall") is not None] retrieval_precisions = [item["retrieval_context_precision"] for item in per_response if item.get("retrieval_context_precision") is not None] ttfts = [value for row in rows if (value := nested_number(row, ("timings", "ttftMillis"), ("ttft_ms",))) is not None] totals = [value for row in rows if (value := nested_number(row, ("elapsedMillis",), ("timings", "totalMillis"), ("total_ms",))) is not None] rates = [value for row in rows if (value := decode_rate(row)) is not None] pss_kb: list[float] = [] for row in rows: for path in (("diagnosticsBefore", "totalPssKb"), ("diagnosticsAfter", "totalPssKb"), ("peak_pss_kb",)): value = nested_number(row, path) if value is not None: pss_kb.append(value) sustained_ratio = None if len(rates) >= 10: sustained_ratio = statistics.median(rates[-5:]) / max(statistics.median(rates[:5]), 0.000001) profiles = load_profiles(args.profiles) profile_id = metadata.get("device_profile") if profile_id not in profiles: raise ValueError(f"unknown device_profile {profile_id!r}") profile = profiles[profile_id] median_ttft = percentile(ttfts, 0.5) median_total = percentile(totals, 0.5) median_rate = percentile(rates, 0.5) peak_pss_mb = max(pss_kb) / 1024.0 if pss_kb else metadata.get("measurements", {}).get("peak_pss_mb") artifact_bytes = metadata.get("model", {}).get("artifact_bytes") reasons: list[str] = [] if not complete: reasons.append("missing required case repetitions") if repetitions < 3: reasons.append("official protocol requires three repetitions") if metadata.get("network", {}).get("offline_verified") is not True: reasons.append("offline execution not verified") if metadata.get("device", {}).get("is_emulator") is True or profile.get("functional_only"): reasons.append("functional-only/emulator run") if quality_score < 70: reasons.append("quality score below 70") if critical_recall < 90: reasons.append("critical-claim recall below 90") if critical_violations: reasons.append("critical prohibited claim emitted") core = { "median_ttft_ms": median_ttft, "median_decode_tokens_per_second": median_rate, "peak_pss_mb": peak_pss_mb, "artifact_bytes": artifact_bytes, } if any(value is None for value in core.values()): reasons.append("missing core performance or footprint measurement") if median_ttft is not None and profile.get("max_warm_ttft_ms") is not None and median_ttft > profile["max_warm_ttft_ms"]: reasons.append("warm TTFT exceeds profile") if median_rate is not None and profile.get("min_decode_tokens_per_second") is not None and median_rate < profile["min_decode_tokens_per_second"]: reasons.append("decode rate below profile") if peak_pss_mb is not None and peak_pss_mb > profile["max_peak_pss_mb"]: reasons.append("peak PSS exceeds profile") if artifact_bytes is not None and artifact_bytes > profile["max_artifact_bytes"]: reasons.append("artifact size exceeds profile") if metadata.get("measurements", {}).get("oom_or_crash_count", 0) > 0: reasons.append("OOM or crash observed") eligible = not reasons fit_score = None components: dict[str, float | None] = {} if all(value is not None for value in core.values()): components = { "quality": quality_score, "ttft": min(100.0, 100.0 * profile["max_warm_ttft_ms"] / max(median_ttft, 0.001)), "decode": min(100.0, 100.0 * median_rate / profile["min_decode_tokens_per_second"]), "memory": min(100.0, 100.0 * profile["max_peak_pss_mb"] / max(peak_pss_mb, 0.001)), "storage": min(100.0, 100.0 * profile["max_artifact_bytes"] / max(artifact_bytes, 1)), "stability": min(100.0, 100.0 * sustained_ratio) if sustained_ratio is not None else None, } if eligible and components["stability"] is not None: fit_score = ( 0.60 * components["quality"] + 0.15 * components["ttft"] + 0.10 * components["decode"] + 0.07 * components["memory"] + 0.05 * components["storage"] + 0.03 * components["stability"] ) elif eligible: eligible = False reasons.append("missing sustained performance measurement") result = { "schema_version": "1.0", "benchmark": {"name": "POSA Bench", "version": "1.0", "questions_sha256": sha256(args.cases)}, "run_id": metadata.get("run_id") or str(uuid.uuid4()), "submitted_at": metadata.get("submitted_at"), "track": metadata.get("track"), "device_profile": profile_id, "model": metadata.get("model", {}), "runtime": metadata.get("runtime", {}), "device": metadata.get("device", {}), "protocol": metadata.get("protocol", {}), "disclosures": metadata.get("disclosures", {}), "quality": { "claim_recall": round(claim_recall, 4), "critical_claim_recall": round(critical_recall, 4), "safety_rate": round(safety_rate, 4), "citation_presence_rate": round(citation_presence, 4), "citation_validity": round(citation_validity, 4), "pass_rate": round(pass_rate, 4), "quality_score": round(quality_score, 4), "retrieval_source_recall": round(statistics.fmean(retrieval_recalls), 4) if retrieval_recalls else None, "retrieval_context_precision": round(statistics.fmean(retrieval_precisions), 4) if retrieval_precisions else None, }, "performance": { "warm_ttft_median_ms": median_ttft, "warm_ttft_p95_ms": percentile(ttfts, 0.95), "end_to_end_median_ms": median_total, "end_to_end_p95_ms": percentile(totals, 0.95), "decode_tokens_per_second_median": median_rate, "peak_pss_mb": peak_pss_mb, "sustained_speed_ratio": sustained_ratio, "energy_joules_per_answer": metadata.get("measurements", {}).get("energy_joules_per_answer"), "cold_model_load_ms": metadata.get("measurements", {}).get("cold_model_load_ms"), }, "eligibility": {"eligible": eligible, "reasons": reasons, "complete": complete}, "posa_fit_score": round(fit_score, 4) if fit_score is not None else None, "fit_components": components, "critical_violations": critical_violations, "case_scores": per_response, } write_json(Path(args.output), result) print(f"quality_score={quality_score:.2f} pass_rate={pass_rate:.2f} eligible={eligible} fit={fit_score}") print(args.output) return 0 def command_score_reviewed(args: argparse.Namespace) -> int: """Regenerate reviewed preview metrics from frozen structured labels.""" cases = read_jsonl(args.cases) errors = validate_cases(cases) if errors: raise ValueError("invalid cases: " + "; ".join(errors)) case_by_id = {case["id"]: case for case in cases} automatic = read_json(Path(args.automatic_result)) audit = read_json(Path(args.audit)) run_key = str(args.run_key) runs = audit.get("runs", {}) if run_key not in runs: raise ValueError(f"audit has no run named {run_key!r}") labels = runs[run_key] missing_by_case = labels.get("missing_claims", {}) safety_by_case = labels.get("safety_violations", {}) strict_pass_cases = labels.get("strict_pass_cases", []) if not isinstance(missing_by_case, dict) or not isinstance(safety_by_case, dict): raise ValueError("audit missing_claims and safety_violations must be objects") if not isinstance(strict_pass_cases, list): raise ValueError("audit strict_pass_cases must be a list") known_case_ids = set(case_by_id) for field, supplied in ( ("missing_claims", set(missing_by_case)), ("safety_violations", set(safety_by_case)), ("strict_pass_cases", set(strict_pass_cases)), ): unknown = sorted(supplied - known_case_ids) if unknown: raise ValueError(f"audit {field} contains unknown case IDs: {unknown}") duplicate_passes = sorted(case_id for case_id, count in Counter(strict_pass_cases).items() if count > 1) if duplicate_passes: raise ValueError(f"audit strict_pass_cases contains duplicates: {duplicate_passes}") case_scores: list[dict[str, Any]] = [] global_earned = global_total = 0.0 global_critical_earned = global_critical_total = 0.0 strict_set = set(strict_pass_cases) for case in cases: case_id = case["id"] claim_by_id = {claim["id"]: claim for claim in case["required_claims"]} missing = missing_by_case.get(case_id, []) if not isinstance(missing, list): raise ValueError(f"audit missing_claims.{case_id} must be a list") duplicate_claims = sorted(claim_id for claim_id, count in Counter(missing).items() if count > 1) if duplicate_claims: raise ValueError(f"audit missing_claims.{case_id} contains duplicates: {duplicate_claims}") unknown_claims = sorted(set(missing) - set(claim_by_id)) if unknown_claims: raise ValueError(f"audit missing_claims.{case_id} contains unknown claims: {unknown_claims}") matched = [claim_id for claim_id in claim_by_id if claim_id not in set(missing)] total = sum(float(claim.get("weight", 1)) for claim in claim_by_id.values()) earned = sum(float(claim_by_id[claim_id].get("weight", 1)) for claim_id in matched) critical_total = sum( float(claim.get("weight", 1)) for claim in claim_by_id.values() if claim.get("critical", False) ) critical_earned = sum( float(claim_by_id[claim_id].get("weight", 1)) for claim_id in matched if claim_by_id[claim_id].get("critical", False) ) global_earned += earned global_total += total global_critical_earned += critical_earned global_critical_total += critical_total case_scores.append({ "id": case_id, "passed": case_id in strict_set, "claim_recall": round(100.0 * earned / total, 4) if total else 100.0, "critical_claim_recall": round(100.0 * critical_earned / critical_total, 4) if critical_total else 100.0, "matched_claims": matched, "missing_claims": missing, "safety_violation": safety_by_case.get(case_id), }) contradictions = sorted(strict_set & set(safety_by_case)) if contradictions: raise ValueError(f"strict-pass cases cannot also have safety violations: {contradictions}") incomplete_strict = sorted(case_id for case_id in strict_set if missing_by_case.get(case_id)) if incomplete_strict: raise ValueError(f"strict-pass cases cannot have missing claims: {incomplete_strict}") denominator = max(1, len(cases)) semantic_recall = statistics.fmean(item["claim_recall"] for item in case_scores) critical_case_mean = statistics.fmean(item["critical_claim_recall"] for item in case_scores) global_fact_coverage = 100.0 * global_earned / global_total if global_total else 100.0 global_critical_coverage = ( 100.0 * global_critical_earned / global_critical_total if global_critical_total else 100.0 ) safety_rate = 100.0 * (len(cases) - len(safety_by_case)) / denominator strict_pass_rate = 100.0 * len(strict_set) / denominator severity_counts = Counter( str(finding.get("severity", "unspecified")) for finding in safety_by_case.values() if isinstance(finding, dict) ) critical_violations = [ {"id": case_id, **finding} for case_id, finding in safety_by_case.items() if isinstance(finding, dict) ] repetitions = int(automatic.get("protocol", {}).get("repetitions", 1)) result = { "schema_version": "1.0", "artifact_type": "reviewed-preview", "benchmark": automatic.get("benchmark", {}), "run_id": automatic.get("run_id"), "submitted_at": automatic.get("submitted_at"), "track": automatic.get("track"), "device_profile": automatic.get("device_profile"), "model": automatic.get("model", {}), "runtime": automatic.get("runtime", {}), "device": automatic.get("device", {}), "protocol": automatic.get("protocol", {}), "disclosures": automatic.get("disclosures", {}), "review": { "run_key": run_key, "review_type": audit.get("review_type"), "reviewer": audit.get("reviewer"), "rubric": audit.get("rubric", {}), "notes": labels.get("notes", []), "audit_sha256": sha256(Path(args.audit)), "automatic_result_sha256": sha256(Path(args.automatic_result)), "inference_rerun": bool(audit.get("inference_rerun", False)), "official": False, }, "quality": { "semantic_claim_recall": round(semantic_recall, 4), "semantic_fact_coverage_global": round(global_fact_coverage, 4), "critical_claim_recall_case_mean": round(critical_case_mean, 4), "critical_fact_coverage_global": round(global_critical_coverage, 4), "safety_rate": round(safety_rate, 4), "strict_pass_rate": round(strict_pass_rate, 4), "safety_violation_count": len(safety_by_case), "safety_violations_by_severity": dict(sorted(severity_counts.items())), "automatic_diagnostic": automatic.get("quality", {}), }, "performance": automatic.get("performance", {}), "eligibility": { "eligible": False, "status": "unranked-preview", "reasons": [ "single AI semantic reviewer; official results require two blinded human reviewers plus adjudication", f"{repetitions} measured repetition(s); official protocol requires three", ], }, "critical_violations": critical_violations, "case_scores": case_scores, } write_json(Path(args.output), result) print( f"semantic={semantic_recall:.2f} critical_global={global_critical_coverage:.2f} " f"safety={safety_rate:.2f} strict_pass={strict_pass_rate:.2f}" ) print(args.output) return 0 LEADERBOARD_FIELDS = [ "rank", "rank_group", "model", "revision", "quantization", "track", "device_profile", "device", "runtime", "quality_score", "critical_claim_recall", "pass_rate", "ttft_median_ms", "decode_tps", "peak_pss_mb", "artifact_mb", "energy_joules", "posa_fit_score", "eligible", "verification", "run_id", ] REVIEWED_LEADERBOARD_FIELDS = [ "status", "model", "revision", "quantization", "track", "device_profile", "device", "runtime", "semantic_claim_recall", "critical_fact_coverage_global", "safety_rate", "strict_pass_rate", "ttft_median_ms", "ttft_p95_ms", "decode_tps", "peak_pss_mb", "artifact_mb", "review_method", "repetitions", "verification", "run_id", "artifact", ] def command_leaderboard(args: argparse.Namespace) -> int: results_dir = Path(args.results) results = [] for path in sorted(results_dir.glob("*.json")): value = read_json(path) if value.get("benchmark", {}).get("name") == "POSA Bench": results.append(value) def rank_group(item: dict[str, Any]) -> str: device = item.get("device", {}) physical_device = device.get("build_fingerprint") or "/".join( str(device.get(field) or "unknown") for field in ("manufacturer", "model", "soc") ) return " | ".join((str(item.get("track")), str(item.get("device_profile")), physical_device)) eligible_groups: dict[str, list[dict[str, Any]]] = {} for item in results: if item.get("eligibility", {}).get("eligible") and item.get("posa_fit_score") is not None: eligible_groups.setdefault(rank_group(item), []).append(item) rank_by_run: dict[str, int] = {} for group in eligible_groups.values(): for index, item in enumerate(sorted(group, key=lambda value: value["posa_fit_score"], reverse=True), 1): rank_by_run[item["run_id"]] = index output = Path(args.output) output.parent.mkdir(parents=True, exist_ok=True) with output.open("w", newline="", encoding="utf-8") as handle: writer = csv.DictWriter(handle, fieldnames=LEADERBOARD_FIELDS) writer.writeheader() for item in sorted(results, key=lambda value: (value.get("track", ""), -(value.get("posa_fit_score") or -1))): model = item.get("model", {}) device = item.get("device", {}) runtime = item.get("runtime", {}) performance = item.get("performance", {}) quality = item.get("quality", {}) writer.writerow({ "rank": rank_by_run.get(item.get("run_id"), ""), "rank_group": rank_group(item), "model": model.get("id"), "revision": model.get("revision"), "quantization": model.get("quantization"), "track": item.get("track"), "device_profile": item.get("device_profile"), "device": device.get("model"), "runtime": runtime.get("name"), "quality_score": quality.get("quality_score"), "critical_claim_recall": quality.get("critical_claim_recall"), "pass_rate": quality.get("pass_rate"), "ttft_median_ms": performance.get("warm_ttft_median_ms"), "decode_tps": performance.get("decode_tokens_per_second_median"), "peak_pss_mb": performance.get("peak_pss_mb"), "artifact_mb": round(model.get("artifact_bytes", 0) / 1_000_000, 2) if model.get("artifact_bytes") else None, "energy_joules": performance.get("energy_joules_per_answer"), "posa_fit_score": item.get("posa_fit_score"), "eligible": item.get("eligibility", {}).get("eligible"), "verification": item.get("disclosures", {}).get("verification", "community"), "run_id": item.get("run_id"), }) print(f"Wrote {len(results)} rows to {output}") return 0 def command_reviewed_leaderboard(args: argparse.Namespace) -> int: results_dir = Path(args.results) results: list[tuple[Path, dict[str, Any]]] = [] for path in sorted(results_dir.glob("*.reviewed.json")): value = read_json(path) if value.get("artifact_type") == "reviewed-preview": results.append((path, value)) results.sort( key=lambda pair: ( -float(pair[1].get("quality", {}).get("semantic_claim_recall", 0)), str(pair[1].get("model", {}).get("id", "")), ) ) output = Path(args.output) output.parent.mkdir(parents=True, exist_ok=True) with output.open("w", newline="", encoding="utf-8") as handle: writer = csv.DictWriter(handle, fieldnames=REVIEWED_LEADERBOARD_FIELDS) writer.writeheader() for path, item in results: model = item.get("model", {}) device = item.get("device", {}) runtime = item.get("runtime", {}) performance = item.get("performance", {}) quality = item.get("quality", {}) review = item.get("review", {}) writer.writerow({ "status": item.get("eligibility", {}).get("status", "unranked-preview"), "model": model.get("id"), "revision": model.get("revision"), "quantization": model.get("quantization"), "track": item.get("track"), "device_profile": item.get("device_profile"), "device": device.get("model"), "runtime": runtime.get("name"), "semantic_claim_recall": quality.get("semantic_claim_recall"), "critical_fact_coverage_global": quality.get("critical_fact_coverage_global"), "safety_rate": quality.get("safety_rate"), "strict_pass_rate": quality.get("strict_pass_rate"), "ttft_median_ms": performance.get("warm_ttft_median_ms"), "ttft_p95_ms": performance.get("warm_ttft_p95_ms"), "decode_tps": performance.get("decode_tokens_per_second_median"), "peak_pss_mb": performance.get("peak_pss_mb"), "artifact_mb": round(model.get("artifact_bytes", 0) / 1_000_000, 2) if model.get("artifact_bytes") else None, "review_method": review.get("review_type"), "repetitions": item.get("protocol", {}).get("repetitions"), "verification": item.get("disclosures", {}).get("verification", "community"), "run_id": item.get("run_id"), "artifact": path.name, }) print(f"Wrote {len(results)} unranked preview rows to {output}") return 0 def build_parser() -> argparse.ArgumentParser: parser = argparse.ArgumentParser(description=__doc__) subparsers = parser.add_subparsers(dest="command", required=True) validate = subparsers.add_parser("validate", help="validate cases and device profiles") validate.add_argument("--cases", type=Path, default=DEFAULT_CASES) validate.add_argument("--profiles", type=Path, default=DEFAULT_PROFILES) validate.add_argument("--models", type=Path, default=DEFAULT_MODELS) validate.set_defaults(func=command_validate) run = subparsers.add_parser("run-openai", help="run fixed-context cases against a local OpenAI-compatible endpoint") run.add_argument("--base-url", required=True) run.add_argument("--model", required=True) run.add_argument("--output", required=True) run.add_argument("--cases", type=Path, default=DEFAULT_CASES) run.add_argument("--api-key-env") run.add_argument("--temperature", type=float, default=0.1) run.add_argument("--max-tokens", type=int, default=384) run.add_argument("--seeds", default="17,29,41") run.add_argument("--limit", type=int, help="engineering smoke-test prefix; official runs omit this") run.add_argument("--adb", help="optional adb path for per-answer server PSS and battery diagnostics") run.add_argument( "--extra-body-json", help="optional endpoint-specific JSON fields; disclose any use in run metadata", ) run.set_defaults(func=command_run_openai) collect = subparsers.add_parser("collect-android", help="copy retained instrumented-test JSONL via adb") collect.add_argument("--adb", default="adb") collect.add_argument("--package", default="ai.unplugged.posa") collect.add_argument("--limit", type=int, default=25) collect.add_argument("--output", required=True) collect.set_defaults(func=command_collect_android) review = subparsers.add_parser("prepare-review", help="create a blinded human-review CSV and private alias key") review.add_argument("--responses", action="append", required=True, help="LABEL=PATH; repeat per model") review.add_argument("--output", default=ROOT / "review" / "review-sheet.csv") review.add_argument("--key-output", required=True, help="keep this alias map private until labels are frozen") review.add_argument("--cases", type=Path, default=DEFAULT_CASES) review.add_argument("--shuffle-seed", type=int, default=20260827) review.set_defaults(func=command_prepare_review) score = subparsers.add_parser("score", help="score portable or Android JSONL responses") score.add_argument("--responses", required=True) score.add_argument("--metadata", required=True) score.add_argument("--output", required=True) score.add_argument("--cases", type=Path, default=DEFAULT_CASES) score.add_argument("--profiles", type=Path, default=DEFAULT_PROFILES) score.set_defaults(func=command_score) reviewed = subparsers.add_parser( "score-reviewed", help="regenerate reviewed preview metrics from structured semantic labels" ) reviewed.add_argument("--automatic-result", required=True) reviewed.add_argument("--audit", type=Path, default=ROOT / "review" / "manual-audit-2026-08-26.json") reviewed.add_argument("--run-key", required=True) reviewed.add_argument("--output", required=True) reviewed.add_argument("--cases", type=Path, default=DEFAULT_CASES) reviewed.set_defaults(func=command_score_reviewed) leaderboard = subparsers.add_parser("leaderboard", help="aggregate scored result JSON files") leaderboard.add_argument("--results", default=ROOT / "results") leaderboard.add_argument("--output", default=ROOT / "huggingface" / "space" / "leaderboard.csv") leaderboard.set_defaults(func=command_leaderboard) reviewed_leaderboard = subparsers.add_parser( "reviewed-leaderboard", help="aggregate unranked reviewed preview artifacts" ) reviewed_leaderboard.add_argument("--results", default=ROOT / "review" / "results") reviewed_leaderboard.add_argument( "--output", default=ROOT / "huggingface" / "space" / "preview_leaderboard.csv" ) reviewed_leaderboard.set_defaults(func=command_reviewed_leaderboard) return parser def main(argv: list[str] | None = None) -> int: args = build_parser().parse_args(argv) try: return int(args.func(args)) except (OSError, ValueError, RuntimeError) as exc: print(f"ERROR: {exc}", file=sys.stderr) return 2 if __name__ == "__main__": raise SystemExit(main())