scripts/build_list.py

"""Stage 1: build the 1930s banned list and upload it to the destination repo.



Seed = croqaz/vintage-ft-v1 banned.txt (D1, 1900 cutoff) embedded verbatim + a

supplement of post-1930 terms. ALLOW_1930 strips the pre-1931-legit entries for

the 1930 cutoff; EXTRA_1930_BANNED backfills clear post-1930 terms. The final list,

allow-list, removed-from-seed audit, and metadata are written under _banned/ in

{DST_REPO} and the total term count is printed.



Run:  python scripts/build_list.py

"""
import json
import re
import time

from huggingface_hub import CommitOperationAdd, hf_hub_download

import config
from common import api, HF_TOKEN, ensure_dst_repo


def _norm(term):
    """Normalize a term for matching: lowercase, collapse internal whitespace."""
    return re.sub(r"\s+", " ", term.strip().lower())


def _clean_terms(raw_iterable):
    """Drop blanks/comments/code-fence markers; normalize; dedupe (keep order)."""
    seen = set()
    out = []
    for line in raw_iterable:
        s = line.strip()
        if not s or s.startswith("#") or s.startswith("```"):
            continue
        t = _norm(s)
        if t and t not in seen:
            seen.add(t)
            out.append(t)
    return out


# --- SEED: post-1900 anachronisms. This is croqaz/vintage-ft-v1 banned.txt
#     (the D1 "hard veto" list, built for a 1900 cutoff) embedded VERBATIM, plus
#     a supplementary section of post-1930 terms croqaz lacks. It intentionally
#     still contains pre-1931-legit terms (aeroplane, radio, einstein, quantum,
#     x-ray, ...); ALLOW_1930 removes them below, and _banned/removed_from_seed.json
#     records exactly what was stripped as the audit of the 1900 -> 1930 shift.
#     Embedded (not fetched) so the notebook is reproducible and offline-safe.


SEED_RAW = """

# Curated post-1900 anachronisms -> HARD VETO (D1).

# A hit forces the text to be dropped. Terms here are AUTHORITATIVE: they are

# never overridden by the allow-list (unlike the auto-parsed banned_seed.txt).

# Phrases match whole-word, case-insensitive; a short phrase ("world war") also

# matches longer mentions ("world war ii").



1000W of power

100W of power

180W of power

2000W of power

20th century fox

21st century

9/11

abortion law

Ada program

aeroplane

ai language model

aircraft

airliner

airplane

airport

always-on display

android phone

antibiotic

antimatter

apartheid

app store

arcade games

artificial intelligence

as an AI

astronaut

atomic bomb

automation

bash script

berlin wall

big bang

big-bang

binary pulsar

bioacoustic

black hole

blitzkrieg

blockchain

blog

blog post

blogger

bluetooth

boeing 777

bubble gum

bubblegum

C program

C#

c++

C++

cell phone

cell tower

cellphone

cellular connectivity

chatbot

chatgpt

chemotherapy

chromosome

cinema

climate change

Clojure function

Clojure program

code snippet

cold war

command-line

compiler

computer

COVID-19

CPU

credit card

credit-card

CSS

CSV file

d-day

dark energy

dark matter

database

DDR3

DDR4

DDR5

DDR6

DDR7

deep learning

desktop computer

digital edition

digital media

diode

DNA

download

drone

DVD

DynamoDB

e-book

e-commerce

e-learning

e-mail

ebook

effective field theory

electric beard trimmer

electric toothbrush

electro-optic

Elixir function

Elixir program

email

emoji

empire state building

EU countries

European Union

eurovision

Excel spreadsheet

Excel table

fascism

fascist

fax

ferromagnetic material

ferromagnetism

fiber-optic

fibre-optic

final fantasy X

first world war

FizzBuzz

fortran

Frenkel defect

FTP server

function in C

function in Go

function in R

gaming

gaming industry

gender affirming

gender nonconforming

gene expression

gene therapy

general relativity

genome

gestapo

gigabyte

Go code

Go function

Go program

godzilla

Golang function

Golang microservice

Golang program

GPS

GPU

graphene

gravitational redshift

great depression

Grignard reagent

hardware

harry potter

hashtag

haskell code

haskell function

haskell program

Hawking radiation

HDD

headphone

helicopter

Higgs boson

Higgs field

hip hop

hip-hop

hiroshima

hollywood

holocaust

hormone

HTML

http

HTTP POST

Hubble's Law

hydrogen bomb

illegal immigratio

internet

ionic compounds

iOS application

iPad

iPhone

iron curtain

iWatch

java applet

java class

java code

java function

java program

JavaScript

jazz

jet engine

jet flight

jetliner

korean war

kotlin function

kotlin program

laptop

Large Hadron Collider

large language model

laser

Lattice QCD

LCD

league of nations

lepton-jet

leptoquark

lesbian

lgbt

Liquid-crystal display

low bass

machine learning

matlab

mechanisation

mechanization

megabyte

metal-organic

metamaterial

Michelson-Morley experiment

microchip

microprocessor

microservice

microwave

minecraft

mobile phone

molecule

moon landing

motherboard

motion picture

movie

MQTT broker

mRNA

MySQL

NAFTA

navigation system

Navy Seals

nazi

nazism

neural network

neurology

neutron

Node.js

nuclear bomb

nuclear reaction

nuclear reactor

nuclear weapon

nuremberg trials

nylon

OCaml function

official website

omega-3

online

optical fiber

PDF version

pearl harbor

penicillin

perl function

perl program

perl script

PHP

pink floyd

pixel

plastic-free

plutonium

podcast

polyester

polyethylene

polymer

polystyrene

pornographic

PostgreSQL

PowerShell

program in C++

program in Javascript

program in Python

programming language

proton

python code

Python Flask

python function

python program

python script

pytorch

quantum

quantum chromodynamics

quantum field theory

quark-gluon plasma

R code

radar

radio broadcast

reality television

reality TV

reboot

Redis key-value

regular expression

REST API

RESTful API

rna

rock 'n' roll

rock and roll

rocket ship

Ruby code

Ruby function

Ruby program

Ruby script

Rust function

Rust program

same-sex marriage

satellite

satellite-dish

Scala code

Scala program

Schottky defect

screenplay

SDK

search engine

search-engine

second world war

selfie

semiconductor

server in Java

shamebloody

Shell command

smart phone

smart watch

smartphone

smartwatch

social media

software

sonar

soviet union

space race

space shuttle

space station

spacecraft

spaceflight

spaceship

special effects

spectral classification

spectrometry

spin electronics

spintronics

SQL

SSD

star wars

stellar classification

stereo

Super Bowl

super mario

superconductivity

superconductor

swift app

swift function

synthetic tissue

television

terrabyte

the beatles

the movies

trans activist

trans people

transistor

tRNA

trotsky

try-catch block

TV show

twentieth century

ultra lightweight

TypeScript

united nations

upload

ussr

vegan

veganism

video camera

video game

videogame

vietnam war

virtual reality

vitamin

web site

website

WebSocket

went viral

wi-fi

wifi

Wittig reagent

world war

WW1

WW2

WW3

wwi

wwii

www

X-files

X-men

X-ray astronomy

xanthan gum



# --- people ---

adolf hitler

albert einstein

anne sexton

benito mussolini

bob dylan

bob marley

carlos santana

donald trump

einstein

elvis presley

ernest hemingway

franklin roosevelt

franz kafka

gandhi

hitler

j. d. salinger

j. r. r. tolkien

j.d. salinger

j.r.r. tolkien

James Bond

john f. kennedy

john fitzgerald kennedy

joseph stalin

lenin

mahatma gandhi

mao zedong

marilyn monroe

martin luther king jr.

mussolini

Nikola Tesla

Robert Downey Jr.

stalin

sylvia plath

taylor swift

Tony Stark

vladimir lenin

walt disney

william faulkner



# --- brands (founded post-1900) ---

AstraZeneca

burger king

facebook

github

google

instagram

mcdonald's

mcdonalds

microsoft

netflix

nintendo

openai

panasonic

samsung

tiktok

twitter

youtube

# --- supplementary post-1930 terms (not in croqaz; added for 1930 coverage) ---

positron

deuterium

cyclotron

electron microscope

nuclear fission

nuclear fusion

manhattan project

chain reaction

radioactive fallout

carbon dating

radiocarbon

standard model

quantum entanglement

quantum computer

double helix

genetic engineering

stem cell

crispr

supersonic

cosmonaut

integrated circuit

semiconductor chip

silicon chip

personal computer

operating system

source code

web browser

web page

world wide web

usb

hyperlink

mobile app

transistor radio

world war ii

world war 2

nazi germany

nagasaki

new deal

isbn

paperback edition

project gutenberg

xerox

atom bomb

maser

jet aircraft

jet plane

nasa

nato

fbi

cia

wikipedia

bitcoin

cryptocurrency

coca-cola company

disney world

kleenex

"""

# --- ALLOW_1930: terms LEGITIMATE at a 1930 cutoff -> removed from the seed.
#     Each is attested in common use by 1930 or earlier. Whole-word, case-insensitive
#     matching means the base term also spares its longer mentions. Conservative:
#     when a term straddles ~1900-1935, ALLOW it (the log-prior filter is the backstop).
#     Deliberately NOT allowed (stay banned): neutron(1932), positron(1932),
#     deuterium(1931), empire state building(1931), franklin roosevelt(1933+ assoc).
ALLOW_1930 = {
    # aviation & transport (Wright 1903; airlines/airports common by the 1920s)
    "aeroplane", "airplane", "aircraft", "airliner", "airport", "aviation",
    "aviator", "airship", "zeppelin", "dirigible", "biplane", "monoplane",
    "automobile", "motorcar", "motor car", "motorcycle", "locomotive",
    "steamship", "submarine", "helicopter",
    # comms & media (all household terms by 1930)
    "radio", "radio broadcast", "wireless", "telephone", "telegraph", "telegram",
    "phonograph", "gramophone", "cinema", "cinematograph", "motion picture",
    "moving picture", "movie", "movies", "the movies", "film", "photograph",
    "photography", "camera", "video camera", "typewriter", "stereo", "microphone",
    "television",  # word coined 1900s; per conservative policy, allow
    # physics / chemistry attested <=1930 (Einstein 1905/15; QM 1925-27)
    "quantum", "quantum theory", "quantum mechanics", "quantum physics",
    "relativity", "theory of relativity", "special relativity", "general relativity",
    "photon", "electron", "proton", "atom", "atomic", "atomic theory", "isotope",
    "ion", "radioactivity", "radioactive", "radium", "uranium", "x-ray", "x ray",
    "xray", "roentgen", "spectroscopy", "spectrometry", "spectral classification",
    "stellar classification", "thermodynamics", "entropy", "electromagnetic",
    "ferromagnetism", "ferromagnetic material", "molecule", "ionic compounds",
    "michelson-morley experiment", "cathode ray", "vacuum tube", "polymer",
    "planck", "bohr", "rutherford", "curie",
    # biology / medicine <=1930
    "vitamin", "hormone", "insulin", "gene", "genetics", "chromosome", "heredity",
    "evolution", "natural selection", "darwin", "mendel", "bacteria", "bacterium",
    "virus", "vaccine", "vaccination", "antiseptic", "anesthesia", "anaesthesia",
    "neurology", "germ theory",
    # psychology / social <=1930
    "psychoanalysis", "freud", "subconscious", "unconscious", "psychology",
    # general modern-sounding but <=1930
    "electricity", "electric", "dynamo", "generator", "turbine", "engine",
    "petroleum", "gasoline", "petrol", "kerosene", "celluloid", "rubber",
    "skyscraper", "elevator", "escalator", "subway", "tram", "streetcar",
    "jazz", "ragtime", "aspirin", "plastic", "bakelite", "rayon",
    "mechanization", "mechanisation", "fascism", "fascist",
    # WWI-era history (WWI itself is pre-cutoff; WWII terms are added separately)
    "first world war", "world war", "the world war", "wwi", "ww1", "great war",
    "twentieth century", "nineteenth century", "league of nations",
    # people alive/active & famous by 1930
    "einstein", "albert einstein", "nikola tesla", "lenin", "vladimir lenin",
    "gandhi", "mahatma gandhi", "franz kafka", "leon trotsky", "trotsky",
    "joseph stalin", "stalin", "ernest hemingway", "william faulkner", "walt disney",
}
ALLOW_1930 = {_norm(t) for t in ALLOW_1930}

# --- EXTRA_1930_BANNED: unambiguously post-1930 terms to guarantee coverage
#     even if the seed missed them. Conservative: borderline items (penicillin,
#     antibiotic ~1940s clinical use) are included since they are clearly
#     post-1930 as common terms; if they over-fire, move them to ALLOW_1930. ---
EXTRA_1930_BANNED = {
    # physics/science firmly after 1930
    "neutron",              # Chadwick, 1932
    "positron",             # 1932
    "deuterium",            # 1931
    "nuclear fission",      # 1938
    "nuclear fusion",
    "nuclear reactor", "nuclear physics", "nuclear energy", "nuclear power",
    "atomic bomb", "atom bomb", "hydrogen bomb", "nuclear weapon",
    "manhattan project", "chain reaction",
    "antibiotic", "penicillin", "sulfa drug", "sulfanilamide",
    "dna", "rna", "double helix", "genome", "genetic code",
    "radar", "sonar", "jet engine", "jet aircraft", "jet plane",
    "television broadcast", "transistor",
    "cyclotron",            # 1932 (Lawrence)
    "electron microscope",  # 1931-33
    # computing / digital (all post-1930)
    "computer program", "digital computer", "electronic computer",
    "microprocessor", "microchip", "integrated circuit", "semiconductor",
    "internet", "world wide web", "website", "email", "e-mail", "software",
    "smartphone", "laser", "maser",
    # geopolitics / history post-1930
    "world war ii", "world war 2", "second world war", "cold war",
    "united nations", "nato", "nazi", "nazi germany", "holocaust", "auschwitz",
    "pearl harbor", "hiroshima", "nagasaki", "d-day", "normandy landing",
    "iron curtain", "berlin wall", "soviet union", "ussr",
    "spanish civil war",    # 1936
    # named orgs/entities post-1930
    "nasa", "fbi", "cia", "gestapo", "luftwaffe", "wehrmacht",
    # publishing / format
    "isbn", "paperback", "photocopy", "xerox", "ballpoint",
}
EXTRA_1930_BANNED = {_norm(t) for t in EXTRA_1930_BANNED}


FORMAT_TELL_PATTERNS = {
    "copyright_post_1930": r"(?:©|copyright|\(c\))\s*(?:19[3-9]\d|20\d\d)",
    "modern_year_reserved": r"all\s+rights\s+reserved",
    "isbn": r"\bisbn(?:-1[03])?\b",
    "url_www": r"\bwww\.",
    "url_http": r"https?://",
    "url_dotcom": r"\b[\w-]+\.(?:com|org|net|edu|gov|io)\b",
    "loc_cip": r"library\s+of\s+congress\s+cata-?\s*loging|cataloging-in-publication",
    "printed_usa_modern": r"printed\s+in\s+the\s+united\s+states\s+of\s+america",
    "gutenberg_license": r"project\s+gutenberg(?:-tm)?(?:\s+(?:license|ebook|literary))",
    "email_addr": r"\b[\w.-]+@[\w.-]+\.\w{2,}\b",
}

# ---------------------------------------------------------------------------
# TIER SYSTEM -- drop only on strong evidence, to kill polysemy false positives.
#
#   tier 1 (hard)  : one hit drops the doc. Coined well after 1930, no earlier
#                    sense; proper nouns of post-1930 people/brands/products.
#   tier 2 (mod)   : default for real anachronisms. Contributes to a drop but is
#                    not decisive alone (see decide_drop in filter_lib).
#   tier 3 (weak)  : polysemous / has a legitimate pre-1930 meaning ("compiler of
#                    this volume", bee "drone", birdsong "twitter", Black Hole of
#                    Calcutta). NEVER fires alone -- only corroborates.
#   strip  (bpl)   : reproduction / boilerplate artifact (URLs, "all rights
#                    reserved", "photocopy", edition notices). Never drops a doc;
#                    logged so the wrapper can be audited/stripped later.
#
# Assignment = category defaults are TIER 2; the explicit sets below override.
# ---------------------------------------------------------------------------

# Words with a real pre-1930 sense -> must never drop a document on their own.
TIER3_TERMS = {
    # --- from observed false positives ---
    "compiler", "hardware", "gaming", "gaming industry", "drone",
    "satellite", "satellite-dish", "twitter", "black hole", "deep learning",
    "holocaust", "pearl harbor", "great depression", "lesbian",
    # --- other clear polysemy / pre-1930 senses ---
    "computer", "automation", "online", "download", "upload", "reboot", "pixel",
    "diode", "microwave", "fax", "stereo", "antimatter", "supersonic",
    "chemotherapy", "omega-3", "low bass", "ultra lightweight",
    "special effects", "screenplay", "navigation system", "hollywood",
    "hip hop", "hip-hop", "rock and roll", "rock 'n' roll", "climate change",
    # 1931-32 physics: genuinely post-1930 but legitimately discussable in a
    # period text -> require corroboration rather than dropping alone.
    "positron", "neutron", "deuterium", "cyclotron", "antimatter",
    "plastic-free",
}

# Reproduction / boilerplate word-terms -> strip-only (never contribute to drop).
# (The FORMAT_TELL_PATTERNS regexes are treated as strip-only in code as well.)
STRIP_ONLY_TERMS = {
    "photocopy", "paperback", "paperback edition", "pdf version",
    "digital edition", "digital media", "e-book", "ebook",
    "official website", "web site", "website", "web page", "web browser",
    "project gutenberg", "www", "http", "online",
}

# Hard, unambiguous post-1930 coinages + proper nouns -> one hit drops.
TIER1_TERMS = {
    # modern brands / products / platforms
    "chatgpt", "openai", "github", "google", "facebook", "instagram", "tiktok",
    "youtube", "netflix", "microsoft", "samsung", "panasonic", "nintendo",
    "astrazeneca", "burger king", "mcdonald's", "mcdonalds", "disney world",
    "kleenex", "boeing 777", "coca-cola company", "iphone", "ipad", "iwatch",
    "android phone", "app store", "smartphone", "smart phone", "smartwatch",
    "smart watch", "blockchain", "bitcoin", "cryptocurrency", "covid-19",
    "pytorch", "node.js", "minecraft", "super mario", "final fantasy x",
    "godzilla", "star wars", "x-men", "x-files", "harry potter", "super bowl",
    "eurovision",
    # post-1930 named people
    "adolf hitler", "benito mussolini", "mussolini", "mao zedong",
    "donald trump", "taylor swift", "elvis presley", "bob dylan", "bob marley",
    "the beatles", "pink floyd", "marilyn monroe", "john f. kennedy",
    "john fitzgerald kennedy", "martin luther king", "robert downey jr.",
    "tony stark", "james bond", "carlos santana", "anne sexton", "sylvia plath",
    "j.r.r. tolkien", "j. r. r. tolkien", "j.d. salinger", "j. d. salinger",
    # unambiguous post-1930 tech / computing / science
    "transistor", "microprocessor", "microchip", "integrated circuit",
    "semiconductor chip", "silicon chip", "internet", "world wide web",
    "css", "html", "javascript", "typescript", "php", "fortran",
    "dna", "rna", "mrna", "trna", "crispr", "genome", "double helix",
    "stem cell", "gene therapy", "genetic engineering", "nylon",
    "wifi", "wi-fi", "bluetooth", "usb", "gpu", "cpu", "ssd", "hdd", "dvd",
    "laptop", "quantum computer", "quantum chromodynamics",
    "quantum field theory", "large hadron collider", "higgs boson",
    "higgs field", "hawking radiation", "dark matter", "dark energy",
    "graphene", "spintronics", "metamaterial",
    "space shuttle", "space station", "moon landing", "spaceflight",
    "spacecraft", "cosmonaut", "astronaut", "nuclear weapon", "nuclear reactor",
    "hydrogen bomb", "atomic bomb", "atom bomb", "manhattan project",
    "nuremberg trials", "auschwitz", "berlin wall", "iron curtain", "cold war",
    "world war ii", "world war 2", "second world war", "wwii", "ww2", "d-day",
}


def assign_tiers(final_terms):
    """Map every banned term to a tier: 1, 2, 3, or 'strip'.



    Precedence: strip-only > tier 3 > tier 1 > default tier 2. (Strip and tier-3

    win over tier-1 so an ambiguous/boilerplate term is never treated as hard.)

    Returns (tiers_by_term, counts).

    """
    strip = {_norm(t) for t in STRIP_ONLY_TERMS}
    t3 = {_norm(t) for t in TIER3_TERMS} - strip
    t1 = {_norm(t) for t in TIER1_TERMS} - strip - t3

    tiers = {}
    for t in final_terms:
        if t in strip:
            tiers[t] = "strip"
        elif t in t3:
            tiers[t] = 3
        elif t in t1:
            tiers[t] = 1
        else:
            tiers[t] = 2
    counts = {
        "tier1": sum(1 for v in tiers.values() if v == 1),
        "tier2": sum(1 for v in tiers.values() if v == 2),
        "tier3": sum(1 for v in tiers.values() if v == 3),
        "strip": sum(1 for v in tiers.values() if v == "strip"),
    }
    return tiers, counts


def build_banned_list():
    """Construct the final banned term list from seed + allow + extra. Returns

    (final_terms_sorted, meta_dict, removed_from_seed_list)."""
    seed = _clean_terms(SEED_RAW.splitlines())
    seed_set = set(seed)

    removed_from_seed = sorted(seed_set & ALLOW_1930)
    kept_from_seed = seed_set - ALLOW_1930

    # Guard: never let an allow-listed term sneak back in via EXTRA.
    extra_effective = EXTRA_1930_BANNED - ALLOW_1930
    added_by_extra = sorted(extra_effective - kept_from_seed)

    final = sorted(kept_from_seed | extra_effective)
    tiers, tier_counts = assign_tiers(final)

    meta = {
        "cutoff_year": config.CUTOFF_YEAR,
        "n_seed": len(seed),
        "n_removed_by_allow": len(removed_from_seed),
        "n_added_by_extra": len(added_by_extra),
        "n_allow_list": len(ALLOW_1930),
        "n_format_tells": len(FORMAT_TELL_PATTERNS),
        "n_final_terms": len(final),
        "n_total_banned_signals": len(final) + len(FORMAT_TELL_PATTERNS),
        "tier_counts": tier_counts,
        "drop_rule": "1x tier1  OR  (>=2 distinct tier2/tier3 with >=1 tier2). "
                     "tier3 never fires alone; strip-only never contributes.",
        "built_at": time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime()),
        "policy": "tiered_confidence",
    }
    return final, meta, removed_from_seed, added_by_extra, tiers


# Terms that contain punctuation the word-tokenizer would split (9/11, c++, ...).


def try_download_banned_list():
    """Load a cached banned list + tiers from the destination repo if present.



    Requires tiers.json to be present; a pre-tier cached list is treated as a

    cache miss so it gets rebuilt with tiers.

    """
    try:
        list_path = hf_hub_download(config.DST_REPO, "_banned/banned_list.txt", repo_type="dataset", token=HF_TOKEN)
        meta_path = hf_hub_download(config.DST_REPO, "_banned/list_meta.json", repo_type="dataset", token=HF_TOKEN)
        tier_path = hf_hub_download(config.DST_REPO, "_banned/tiers.json", repo_type="dataset", token=HF_TOKEN)
        with open(list_path, "r", encoding="utf-8") as f:
            terms = [_norm(l) for l in f if l.strip()]
        with open(meta_path, "r", encoding="utf-8") as f:
            meta = json.load(f)
        with open(tier_path, "r", encoding="utf-8") as f:
            raw = json.load(f)
        # tiers.json stores ints as JSON strings for "strip"; normalize back.
        tiers = {k: (v if v == "strip" else int(v)) for k, v in raw.items()}
        print("Loaded cached banned list + tiers from destination repo.")
        return terms, meta, tiers
    except Exception:
        return None, None, None


def _write_tier_files(d, tiers):
    """Write per-tier text files + tiers.json to local dir d; return their paths."""
    buckets = {1: [], 2: [], 3: [], "strip": []}
    for term, tier in tiers.items():
        buckets[tier].append(term)
    paths = {}
    for tier, name in [(1, "tier1.txt"), (2, "tier2.txt"), (3, "tier3.txt"), ("strip", "strip_only.txt")]:
        p = d / name
        p.write_text("\n".join(sorted(buckets[tier])) + "\n", encoding="utf-8")
        paths[name] = p
    tj = d / "tiers.json"
    with open(tj, "w", encoding="utf-8") as f:
        json.dump({k: v for k, v in sorted(tiers.items())}, f, indent=2)
    paths["tiers.json"] = tj
    return paths


def build_and_upload_banned_list():
    final, meta, removed_from_seed, added_by_extra, tiers = build_banned_list()
    config.ensure_dirs()
    d = config.BANNED_DIR
    (d / "banned_list.txt").write_text("\n".join(final) + "\n", encoding="utf-8")
    (d / "allow_list.txt").write_text("\n".join(sorted(ALLOW_1930)) + "\n", encoding="utf-8")
    with open(d / "removed_from_seed.json", "w", encoding="utf-8") as f:
        json.dump({"removed_from_seed_because_pre_1931": removed_from_seed,
                   "added_by_extra_1930_banned": added_by_extra}, f, indent=2, sort_keys=True)
    with open(d / "list_meta.json", "w", encoding="utf-8") as f:
        json.dump(meta, f, indent=2, sort_keys=True)
    with open(d / "format_tells.json", "w", encoding="utf-8") as f:
        json.dump(FORMAT_TELL_PATTERNS, f, indent=2, sort_keys=True)
    tier_paths = _write_tier_files(d, tiers)

    ops = [
        CommitOperationAdd("_banned/banned_list.txt", str(d / "banned_list.txt")),
        CommitOperationAdd("_banned/allow_list.txt", str(d / "allow_list.txt")),
        CommitOperationAdd("_banned/removed_from_seed.json", str(d / "removed_from_seed.json")),
        CommitOperationAdd("_banned/list_meta.json", str(d / "list_meta.json")),
        CommitOperationAdd("_banned/format_tells.json", str(d / "format_tells.json")),
    ]
    ops += [CommitOperationAdd(f"_banned/{name}", str(p)) for name, p in tier_paths.items()]

    api.create_commit(
        repo_id=config.DST_REPO, repo_type="dataset", operations=ops,
        commit_message=(f"add 1930s tiered banned list ({meta['n_final_terms']} terms; "
                        f"tiers {meta['tier_counts']})"),
    )
    return final, meta, tiers


def load_or_build():
    """Return (banned_terms, meta, tiers): cached copy unless FORCE_REBUILD_LIST."""
    ensure_dst_repo()
    if config.FORCE_REBUILD_LIST:
        terms, meta, tiers = None, None, None
    else:
        terms, meta, tiers = try_download_banned_list()
    if terms is None:
        terms, meta, tiers = build_and_upload_banned_list()
    return terms, meta, tiers


def main():
    terms, meta, tiers = load_or_build()
    print("=" * 60)
    print("BANNED LIST SUMMARY (tiered)")
    print("=" * 60)
    print(json.dumps(meta, indent=2, sort_keys=True))
    tc = meta.get("tier_counts", {})
    print(f"\nTOTAL WORDS/PHRASES IN BANNED LIST: {len(terms):,}")
    print(f"  tier 1 (hard, 1 hit drops):        {tc.get('tier1', 0):,}")
    print(f"  tier 2 (moderate):                 {tc.get('tier2', 0):,}")
    print(f"  tier 3 (weak, never fires alone):  {tc.get('tier3', 0):,}")
    print(f"  strip-only (boilerplate, no drop): {tc.get('strip', 0):,}")
    print(f"Plus {len(FORMAT_TELL_PATTERNS)} format-tell patterns (strip-only).")
    return terms, meta, tiers


if __name__ == "__main__":
    main()