#!/usr/bin/env python3 """ PFLT / FSOT communicator benchmarks (NOT LLM leaderboard metrics). We intentionally do **not** optimize for: - MMLU / chat preference / next-token perplexity as primary goals We DO report metrics that match this architecture: 1) Lexical grounding accuracy (historical + classical + hieroglyph closed-set) 2) FSOT scalar health (finite S, domain routing, quirk_mod active when observed) 3) Vision contract readiness (stub: labels → meanings map rate) 4) Optional: compare against a thin neural baseline later (same tasks) This is how you stack against cutting-edge AI without becoming an LLM: - Same *task families* where useful (retrieval@1, F1, CER for OCR) - Different *objective* (FSOT-gated interlingua communicator) """ from __future__ import annotations import json import sys from datetime import datetime, timezone from pathlib import Path from typing import Any, Dict, List sys.path.insert(0, str(Path(__file__).resolve().parent)) from PFLT_FSOT_2_1_aligned import PFLT, compute_S_D_chaotic, DOMAIN_PARAMS from vision_stub import vision_translate def bench_classical(pflt: PFLT, n: int = 100) -> Dict[str, Any]: path = Path(__file__).resolve().parent / "data" / "classical_grc_la_promoted_tierA.jsonl" if not path.exists(): return {"ok": False, "error": "missing classical gold"} rows = [] for line in path.read_text(encoding="utf-8").splitlines(): if line.strip(): rows.append(json.loads(line)) rows = rows[:n] hits = 0 for r in rows: out = pflt.translate(r["source_word"], context="historical", target_lang="english") blob = " ".join(out["meanings"]).lower() + " " + out["translation"].lower() gold = r["target_word"].lower() mk = r.get("meaning_key", "").lower().replace("_", " ") if gold in blob or mk.replace("_", " ") in blob or r.get("meaning_key", "").lower() in blob: hits += 1 elif any(t in blob for t in gold.split() if len(t) > 2): hits += 0.5 return { "ok": True, "n": len(rows), "exactish_accuracy": hits / max(1, len(rows)), "task": "classical_grc_la_closed_set", } def bench_hieroglyph(pflt: PFLT) -> Dict[str, Any]: codes = ["A1", "D21", "G17", "N5", "S34", "D4", "G5", "I9", "N35", "X1"] hits = 0 details = [] for c in codes: out = pflt.translate(c, context="hieroglyphic") ok = out["exact_map_rate"] >= 1.0 and out["meanings"] and "generic" not in out["meanings"][0] hits += int(ok) details.append({"code": c, "ok": ok, "meaning": out["meanings"], "S": out["fsot_coherence_S"]}) return { "ok": True, "n": len(codes), "map_accuracy": hits / len(codes), "details": details, "task": "hieroglyph_unikemet_closed_set", } def bench_vision_contract() -> Dict[str, Any]: r = vision_translate(gardiner=["A1", "N5", "S34"]) ok = bool(r.hypotheses) and r.pflt.get("exact_map_rate", 0) >= 1.0 return { "ok": ok, "task": "vision_stub_labels_to_meaning", "map_rate": r.pflt.get("exact_map_rate"), "pipeline": r.pipeline, "translation": r.pflt.get("translation"), } def bench_scalar_health() -> Dict[str, Any]: rows = [] for name, p in DOMAIN_PARAMS.items(): panel = compute_S_D_chaotic( D_eff=float(p["D_eff"]), observed=bool(p["observed"]), delta_psi=float(p["delta_psi"]), delta_theta=float(p["delta_theta"]), ) rows.append( { "domain": name, "S": panel.S, "finite": abs(panel.S) < 1e6 and panel.S == panel.S, "observed": panel.observed, "quirk_mod": panel.quirk_mod, } ) return { "ok": all(r["finite"] for r in rows), "task": "fsot_scalar_domain_health", "n_domains": len(rows), "domains": rows, } def positioning_block() -> Dict[str, Any]: return { "product_class": "FSOT interlingua communicator (symbolic + neural student slots)", "not_building": "frontier general LLM / chat assistant", "vs_cutting_edge": { "frontier_llm": { "strength": "open-domain fluency, broad world knowledge, tool use", "weakness_for_your_goal": "opaque params, weak formal guarantees, poor cross-domain physics unity", "overlap": "can share *benchmark task shapes* (retrieval, OCR CER, translation F1)", }, "classical_mt_interlingua": { "strength": "explainable pipelines, domain lexica", "your_edge": "FSOT scalar teacher + multi-domain (DNA, myth, hieroglyph, cosmos) one geometry", }, "vision_ocr_sota": { "strength": "image→text accuracy", "your_use": "U-Net/OCR as *student eyes* only; meaning remains Unikemet/FSOT gated", }, "neuro_symbolic_ai": { "closest_peer_class": True, "your_edge": "seed-derived law + Lean cross-verification + historical curriculum", }, }, "benchmark_axes_we_own": [ "closed_set_lexicon_accuracy (hist/classical/glyph)", "held_out_gap_fill under FSOT gates (future)", "glyph_detection_top1 / CER (when U-Net online)", "scalar_panel_coherence / kill_criteria pass rate", "cross_domain_same_engine (gene + myth + glyph + H0)", ], "benchmark_axes_we_borrow_not_chase": [ "MMLU as primary KPI", "Chatbot Arena ELO", "next-token perplexity on Common Crawl", ], } def main() -> None: pflt = PFLT() report = { "built_utc": datetime.now(timezone.utc).isoformat(), "positioning": positioning_block(), "lexicon_size": len(pflt.pul_terms), "benchmarks": { "scalar_health": bench_scalar_health(), "hieroglyph": bench_hieroglyph(pflt), "classical": bench_classical(pflt), "vision_contract": bench_vision_contract(), }, } out = Path(__file__).resolve().parent / "data" / "benchmark_report.json" out.parent.mkdir(parents=True, exist_ok=True) out.write_text(json.dumps(report, indent=2, ensure_ascii=False), encoding="utf-8") drive = Path(r"D:\training data\pflt_linguistics\00_manifests\benchmark_report.json") drive.parent.mkdir(parents=True, exist_ok=True) drive.write_text(json.dumps(report, indent=2, ensure_ascii=False), encoding="utf-8") b = report["benchmarks"] print("=== PFLT / FSOT communicator benchmark ===") print(f"lexicon_size: {report['lexicon_size']}") print(f"scalar_health: {b['scalar_health']['ok']} domains={b['scalar_health']['n_domains']}") print(f"hieroglyph map_accuracy: {b['hieroglyph'].get('map_accuracy')}") print(f"classical exactish: {b['classical']}") print(f"vision_contract: {b['vision_contract']['ok']} map={b['vision_contract'].get('map_rate')}") print(f"wrote {out}") print(f"wrote {drive}") if __name__ == "__main__": main()