#!/usr/bin/env python3 """Build the coded manifesto dataset (manifest-coded.csv) and aggregations for the 2026 Swedish valmanifest corpus. Applies coding-rubric.md v3 (the long-form extension of the v2 machinery addendum). Corpus: valmanifest/text/*.txt, extracted with `pdftotext -layout` from the publisher PDFs archived in valmanifest/. Provenance: valmanifest-sources.md. Two stages, deliberately separated by how much they can be trusted: Stage 1 (mechanical, reproducible by anyone running this file): a fixed lexicon is matched against each document; hits are segmented into passages; per-party counts are normalised per 1 000 words. Nothing here is a judgement call. Stage 2 (interpretive, single-coder): each passage carries a hand-assigned `frame`, `as_layer` and `own_claim`. These were assigned by reading the passages this script prints. They have NOT been double-coded. See the caveat in coding-rubric.md v3. """ import csv, re, glob, os, collections # ---------------------------------------------------------------- stage 1: lexicon # The lexicon is CORE + INFRA. It deliberately excludes high-frequency ambiguous words — # `data` (everyday sense), `digital` (adjective, mostly "digitala tjänster"), # `teknik`, `plattform` (collides with *valplattform*), `IT`, `upphandling`. # Those are surveyed separately in coding-rubric.md v3 and are not part of the # headline count, because a presence test is only as good as its false-positive rate. CORE = (r"(artificiell intelligens|maskininlärning|\bAI\b|AI-\w*|AI:s|algoritm\w*|" r"språkmodell\w*|datacenter\w*|datacentral\w*|beräkningskraft\w*|" r"beräkningskapacitet\w*|superdator\w*|molntjänst\w*|öppen källkod|" r"öppna vikter|automatiser\w*|digitaliser\w*)") # INFRA — added after an audit of the terms CORE excludes (see coding-rubric.md v3, # "What the first lexicon missed"). CORE alone produced a FALSE NEGATIVE on the # corpus's strongest sovereignty passage, M-04, because Moderaterna write "moln", # not "molntjänst". Infrastructure- and dependency-language is therefore matched too. INFRA = (r"(\bmoln\b|molnet|halvledare|beroende\w* av utländsk|utländsk teknik|" r"säkerhetskänslig\w*|kritiska system|digital infrastruktur|datadelning|" r"cyberskydd\w*|cybersäkerhet\w*|cyberangrepp\w*|cyberhot\w*|" r"cyberförsvar\w*|cyberförmåga\w*)") LEX = f"({CORE}|{INFRA})" # Passage segmentation: consecutive hits closer than GAP characters are one passage. # GAP=250 chosen because the passage count is stable there (45 at 90, 32 at 250, # 31 at 400) — i.e. it sits on the plateau, not on a slope. GAP = 250 PARTIES = ["S", "M", "SD", "C", "V", "KD", "L", "MP"] def load(): docs = {} for f in sorted(glob.glob("valmanifest/text/*.txt")): p = os.path.basename(f).split("-")[0].upper() docs[p] = re.sub(r"\s+", " ", open(f).read()) return docs def passages(text): hits = list(re.finditer(LEX, text, re.I)) out = [] for m in hits: if out and m.start() - out[-1][1] < GAP: out[-1][1] = m.end() else: out.append([m.start(), m.end()]) return hits, out # --------------------------------------------------- stage 2: hand-assigned codes # frame — how the passage frames the machinery. Value set, grounded in the corpus: # competitiveness growth, jobs, "leda", attracting investment or talent # capacity building or hosting technical capacity (compute, centres, data) # dependency reliance on foreign technology named as a vulnerability # labour retraining, skills, education, adaptation of workers # admin-efficiency faster casework, less administration, agencies using AI # regulation rule-making about AI, domestic or international # rights integrity, data protection, right to understand a decision # risk-social harm to children, deepfakes, disinformation, screens # risk-security cyber, hybrid threats, defence # risk-energy electricity or resource cost of the infrastructure # inclusion digital exclusion of citizens # culture copyright, media, public service # toc table-of-contents artefact; excluded from analysis # # as_layer — the Agentic State's twelve layers, as in rubric v2. ';'-separated, or none. # own_claim — 1 if the passage asserts PUBLIC OWNERSHIP OR CONTROL of models, weights, # compute or data infrastructure. This is the variable the whole exercise tests. CODES = { # --- Centerpartiet: the only manifesto with a dedicated AI chapter (1.6) --- "C-01": ("capacity", "9", 0, "transporter och digital infrastruktur knyter ihop landet"), "C-02": ("toc", "none", 0, "TOC: '1.6 AI och ny teknik som skapar fler jobb'"), "C-03": ("competitiveness", "none", 0, "Sverige ska leda i AI, innovation och ny tillväxt"), "C-04": ("labour", "none", 0, "arbetsmarknaden förändras genom AI, digitalisering"), "C-05": ("competitiveness", "2", 0, "bli bäst på tillämpning av AI i privat och offentlig sektor"), "C-06": ("labour", "3", 0, "Lägg om jobb- och utbildningspolitiken för AI-eran"), "C-07": ("capacity", "9;8", 0, "nationellt AI-center som samlar forskning, beräkningskraft och näringsliv"), "C-08": ("rights", "7;8", 0, "människors rätt att förstå beslut som påverkar deras liv"), "C-09": ("labour", "none", 0, "Ge alla AI-kunskap (elever)"), "C-10": ("inclusion", "1", 0, "Ingen ska hamna utanför i digitaliseringen"), "C-11": ("admin-efficiency","2", 0, "digitaliseringen användas för att förenkla, inte skapa krångel"), "C-12": ("admin-efficiency","2;6", 0, "införa AI i stor skala ... snabbspår för AI-upphandling"), "C-13": ("capacity", "9", 0, "investeringar i vägar, järnväg och digital infrastruktur"), "C-14": ("risk-security", "5;10", 0, "digital infrastruktur ska fungera även i kris"), "C-15": ("regulation", "3", 0, "aktiv röst om AI i EU för gemensamma regler"), "C-16": ("risk-security", "5;10", 0, "kunna betala även vid elavbrott, nätstörningar och cyberangrepp"), # --- Kristdemokraterna: one passage in the whole manifesto --- "KD-01": ("risk-social", "none", 0, "Skydda barn från techjättarnas skadliga algoritmer"), # --- Liberalerna --- "L-01": ("risk-social", "none", 0, "när AI är vardag blir källkritik viktigare"), "L-02": ("risk-social", "none", 0, "Digitaliseringen i skolan har gått för långt"), "L-03": ("risk-social", "none", 0, "stoppa kränkande AI-deepfakes"), "L-04": ("dependency", "9;10", 0, "beroende av digital teknik från andra delar av världen gör oss sårbara; Europa ska stärka egen kapacitet inom halvledare, molntjänster, AI"), "L-05": ("competitiveness", "none", 0, "Sverige ska leda teknikrevolutionen"), "L-06": ("labour", "12", 0, "AI-kunskapslyft"), "L-07": ("risk-social", "none", 0, "när algoritmer styr nyhetsflödet ökar polariseringen"), # --- Moderaterna --- "M-01": ("risk-social", "none", 0, "rätt att stoppa AI-genererade nakenbilder"), "M-02": ("risk-security", "10", 0, "byggt upp en starkare svensk cyberförmåga"), "M-03": ("risk-security", "10", 0, "nationellt cyberskydd för Sveriges kommuner, övervakning dygnet runt"), "M-04": ("dependency", "6;9;10",0, "minska beroenden av utländsk teknik i kritiska system genom strategisk upphandling för svensk och europeisk teknik inom säkerhetskänslig IT, moln och drift"), "M-05": ("culture", "none", 0, "AI utmanar affärsmodeller, upphovsrätt och verifiering"), "M-06": ("risk-social", "none", 0, "AI-kunskap och medie- och informationskunskap"), # --- Socialdemokraterna --- "S-01": ("competitiveness", "none", 0, "peka ut tekniker där Sverige ska vara ledande, som AI"), "S-02": ("competitiveness", "none", 0, "investera i bostäder, järnväg, el, digitalisering, forskning"), "S-03": ("admin-efficiency","2", 0, "digitalisering och AI för att korta handläggningstiderna"), "S-04": ("risk-energy", "none", 0, "krav vid energiintensiva etableringar, exempelvis datacenter"), "S-05": ("labour", "12", 0, "Ett svenskt AI program (kompetens, jfr Hem-pc)"), "S-06": ("labour", "none", 0, "med ny teknik - inte minst AI - viktigare med omställning"), "S-07": ("labour", "none", 0, "kursen i artificiell intelligens inställd (folkbildning)"), "S-08": ("admin-efficiency","2;4;8", 0, "AI-tekniken används för att samköra uppgifter"), "S-09": ("risk-security", "10", 0, "informations- och cybersäkerhet i totalförsvaret"), } FRAMES = ["competitiveness","capacity","dependency","labour","admin-efficiency","regulation", "rights","risk-social","risk-security","risk-energy","inclusion","culture"] LAYER_NAMES = {1:"Public Service Design/UX",2:"Government Workflows",3:"Policy- and Rule-Making", 4:"Regulatory Compliance",5:"Crisis Response",6:"Public Procurement", 7:"Agent Governance",8:"Data and Privacy",9:"Tech Stack", 10:"Cyber Security/Resilience",11:"Public Finance",12:"People and Culture"} # ============================================================== build docs = load() rows, doclevel = [], [] for p in PARTIES: t = docs.get(p, "") words = len(t.split()) hits, pas = passages(t) doclevel.append({"party": p, "words": words, "hits": len(hits), "passages": len(pas), "hits_per_1k": round(len(hits)/words*1000, 2) if words else 0.0}) for i, (a, b) in enumerate(pas, 1): pid = f"{p}-{i:02d}" terms = ";".join(sorted({m.group(0).lower() for m in re.finditer(LEX, t[a:b+1], re.I)})) frame, layer, own, anchor = CODES[pid] rows.append({"id": pid, "party": p, "terms": terms, "frame": frame, "as_layer": layer, "own_claim": own, "anchor": anchor, "context": t[max(0, a-160):b+240].strip()}) missing = set(CODES) ^ {r["id"] for r in rows} assert not missing, f"codes out of sync with extraction: {sorted(missing)}" cols = ["id","party","terms","frame","as_layer","own_claim","anchor","context"] with open("manifest-coded.csv","w",newline="") as f: w = csv.DictWriter(f, fieldnames=cols); w.writeheader() for r in rows: w.writerow(r) # ============================================================== report print("=" * 74) print("VALMANIFEST 2026 — MACHINERY CODING (coding-rubric.md v3)") print("=" * 74) print("\nSTAGE 1 — mechanical. Core-lexicon hits, normalised per 1 000 words.\n") print(f" {'party':<7}{'words':>8}{'hits':>7}{'passages':>10}{'per 1k':>9}") for d in sorted(doclevel, key=lambda x: -x["hits_per_1k"]): print(f" {d['party']:<7}{d['words']:>8}{d['hits']:>7}{d['passages']:>10}{d['hits_per_1k']:>9.2f}") tw = sum(d["words"] for d in doclevel); th = sum(d["hits"] for d in doclevel) print(f" {'TOTAL':<7}{tw:>8}{th:>7}{len(rows):>10}{th/tw*1000:>9.2f}") zero = [d["party"] for d in doclevel if d["hits"] == 0] print(f"\n Manifestos with ZERO core-lexicon hits: {', '.join(zero) if zero else 'none'}") print("\n Absent from ALL eight manifestos (0 occurrences in the whole corpus):") for term in ["språkmodell","öppen källkod","öppna vikter","leverantörsberoende", "digital suveränitet","folkbibliotek för modeller","beräkningscentrum"]: n = sum(len(re.findall(term, t, re.I)) for t in docs.values()) if n == 0: print(f" · {term}") print("\nSTAGE 2 — interpretive, single-coder, NOT double-coded.\n") print(" FRAME — how the machinery is framed when it does appear:\n") fc = collections.Counter(r["frame"] for r in rows if r["frame"] != "toc") n_coded = sum(fc.values()) for fr in FRAMES: if fc[fr]: bar = "#" * fc[fr] parties = ",".join(sorted({r["party"] for r in rows if r["frame"] == fr})) print(f" {fr:<17}{fc[fr]:>3} {bar:<12} {parties}") print(f" {'(toc, excluded)':<17}{sum(1 for r in rows if r['frame']=='toc'):>3}") print("\n LAYER — which of the Agentic State's twelve layers the passage touches:\n") lc = collections.Counter() for r in rows: if r["as_layer"] != "none": for L in r["as_layer"].split(";"): lc[int(L)] += 1 for L in range(1, 13): band = "implementation" if L <= 6 else "ENABLEMENT" print(f" L{L:<3}{LAYER_NAMES[L]:<28}{lc[L]:>3} {'#'*lc[L]:<8} {band}") impl, enab = sum(lc[L] for L in range(1,7)), sum(lc[L] for L in range(7,13)) print(f"\n implementation (1-6): {impl} ENABLEMENT (7-12): {enab}" f" no layer: {sum(1 for r in rows if r['as_layer']=='none')}") print("\n OWNERSHIP CLAIM — passages asserting public ownership or control of") print(" models, weights, compute or data infrastructure:\n") own = [r for r in rows if r["own_claim"] == 1] print(f" {len(own)} of {n_coded} coded passages, across {len(PARTIES)} manifestos" f" and {tw:,} words.") if own: for r in own: print(f" {r['id']} {r['party']}: {r['anchor']}") else: print(" The nearest approaches, none of which makes the claim:") for pid in ("C-07", "M-04", "L-04"): r = next(x for x in rows if x["id"] == pid) print(f" {r['id']} ({r['frame']}, L{r['as_layer']}): {r['anchor']}") print()