# Coding rubric — sentiment & framing of Swedish 2026 election posters Version 1.2 · 2026-09-05 (v1.0 variables unchanged; v2 and v3 addenda below) This rubric turns the qualitative read in `poster-analysis.md` into variables that can be applied consistently across the corpus and compared across parties. It is designed so that a second coder — human or model — given the same image and text would produce the same labels. ## Why "sentiment" is defined the way it is here A campaign poster is, by construction, positive about its own party. Classic polarity (positive/negative) therefore carries almost no information. The variation that matters is in **which emotion is solicited**, **what moral language is used**, **who is addressed**, **who (if anyone) is cast as the antagonist**, and **how much of a causal story is told**. The rubric codes those, plus the visual cues that carry affect (colour temperature, faces, type scale). ## Unit of analysis One **unique design**. Alternate resolutions and file formats of the same artwork collapse to one unit (e.g. S `01-ekonomin.jpg` and `01-ekonomin-1920px.jpg` = one unit). A shared slogan rendered as visibly different artworks (different person, pose, or layout) counts as **separate units**, but they carry the same `design_family` tag so family-level and unit-level counts can both be reported. ## Evidence class (carried from the catalogue) - **A** — clean official asset; text and layout analysed confidently. - **B** — official design seen in a press photograph; perspective/obstruction limits detail. - **C** — press-collage only; identification, not fine-grained visual claims. Every coded row records its evidence class. Findings are reported A-only and all-classes so the weaker material never silently drives a conclusion. --- ## Variables ### 0. Identification | Field | Values | |---|---| | `id` | short code, e.g. `S-01`, `M-WC-07` | | `party` | S, V, L, C, M, KD, MP, SD, MED, SjvP, EnadRöst | | `scope` | national / regional / local | | `evidence` | A / B / C | | `design_family` | free text tag shared by slogan-siblings (e.g. `vi-far-saker-gjorda`) | | `slogan_sv` | verbatim headline (square brackets = uncertain) | | `slogan_en` | plain gloss | ### 1. Dominant affect (`affect_primary`, optional `affect_secondary`) The single emotion the design most works to produce in the viewer. Choose from: | Code | Affect | Recognisable by | |---|---|---| | `gravity` | Seriousness / sober authority | formal portrait, muted palette, "på allvar", restraint | | `hope` | Optimism / forward lift | bright light, "vinner", "framåt", smiling candid | | `pride` | National / group pride | flag-coded colour, "Sverige", collective "vi" | | `fear` | Threat / alarm | crime, violence, danger, security nouns | | `anger` | Indignation / grievance | naming a culprit, injustice, "orättvisa", punishment | | `reassurance` | Protection / safety restored | "trygg", "skydda", a promise to shield | | `solidarity` | Recognition / being-on-your-side | everyday scene, "på din sida", listening posture | | `warmth` | Belonging / community calm | shared places, children + elders, greenery | | `resolve` | Managerial competence / "we deliver" | "vi får saker gjorda", results ledger, busy leaders | | `nostalgia` | Loss & restoration of a past | "igen", "återfå", implied golden condition | Rule: pick `affect_primary` first from text+image gestalt; add `affect_secondary` only if a clearly distinct second register is present (e.g. SD: `nostalgia` primary, `reassurance` secondary). ### 2. Temporal stance toward the system (`stance`) Captures the maintenance/restoration/reform/transform axis from the analysis. | Code | Meaning | |---|---| | `restore` | return to a lost/past condition | | `maintain` | keep present institutions, run them more seriously | | `reform` | adjust discrete settings incrementally | | `transform` | change structural ownership/decision arrangements | ### 3. Emotional intensity (`intensity`, 1–5) 1 = calm/measured · 2 = mild · 3 = moderate · 4 = charged · 5 = high-arousal/alarm. Anchors: S text poster = 1; L pastel policy card = 2; V billionaire-tax joke = 3; M crime card = 4; SD policy receipt / "utvisas" language = 5. ### 4. Constructed subject — who is addressed (`subject`) From the five political subjects in the analysis, extended: `adult-nation` · `squeezed-household` · `autonomous-individual` · `local-community` · `bounded-nation` · `crime-victim` · `senior` · `family-parent` · `taxpayer` ### 5. Antagonist / out-group (`antagonist`) `none` or one of: `criminals` · `immigrants-asylum` · `islamists` · `political-opponents` · `profiteers-billionaires` · `waste-bureaucracy` · `developers-highrise`. Record `antagonist_explicit` = yes/no (named on the poster vs. merely implied). ### 6. Moral foundations (`mf_primary`, `mf_all`) Haidt's foundations; code the primary and list all salient: `care` · `fairness` · `loyalty` · `authority` · `sanctity` · `liberty`. Guide: "trygghet/skydda" → care or authority; "frihet" → liberty; "rättvist/klyftor" → fairness; "Sverige/nation/igen" → loyalty; "straff/ansvar/kontroll" → authority; "moské/renhet" → sanctity. ### 7. Causal completeness (`causal_score`, 0–4) Entman frame elements present on the design. +1 each for: names a **problem**; names a **cause**; assigns **blame/agent**; specifies a **remedy/mechanism**. Slogan-only promise with no diagnosis = 0–1; V's care-vs-profit or MP's inequality→safety = 2–3; SD policy receipt = 4. ### 8. Visual affect cues | Field | Values | |---|---| | `color_temp` | warm / neutral / cool / dark | | `subject_of_image` | leader / citizen(s) / place-environment / text-only / object | | `face_prominence` | none / small / medium / dominant | | `type_scale` | measured / large / oversized | ### 9. Rhetorical device (`device`, may list two) `second-person` · `imperative` · `pun-wordplay` · `joke` · `list-receipt` · `slogan-only` · `testimonial` · `comparison-vs-opponent`. --- ## Coding procedure 1. Collapse files to unique designs; assign `id` and `design_family`. 2. Read `slogan_sv` from the cleanest available asset (or the catalogue transcription for B/C). 3. Code text-driven fields (affect, stance, subject, antagonist, moral foundations, causal, device). 4. Code visual fields from the image. 5. Record evidence class and any uncertainty. ## Reliability plan (stated, partially executed) - Full corpus coded once here (single coder = this model, drawing on the Codex catalogue as a second independent description where it exists — that overlap is used as a lightweight cross-check, not as true double-coding). - **To make any published number citable**, a ~30-design subsample should be independently double-coded (second human or a different model with only the rubric + image), and Krippendorff's α reported per variable. α ≥ 0.67 tentative, ≥ 0.80 firm. Categorical variables (affect, subject, moral foundation) will be the hardest and should be reported with their α. ## Sampling caveat (must travel with every result) The archive is **uneven** by construction (see `poster-sources.md`): S/V/L are complete national sets; C national exists as print masters but the richest C material is a single local Sundbyberg series; SD national designs are photo-crops (class B); M is over-represented by 29 web cards that are 16:9 web format, not street posters; KD/MP/M national appear partly only in a press collage (class C); MED is a single hyper-local Danderyd suite. **Per-party affect distributions are shaped by what was collectable, not only by campaign emphasis.** Counts are reported per design *and* flagged by evidence class; cross-party claims lean on class-A material and on within-party pattern, not raw frequency. --- # v2 addendum — machinery coding Added 2026-09-04. Does **not** alter any v1.0 variable or label; it adds two columns to `coded-data.csv`. ## Why Every number this project publishes has a codebook and, since the reliability pass, a measured α — except the one on the front of the website: *no party has a plan for who owns the machines that will administer Sweden.* That claim was rhetorical. This addendum makes it a coded, reproducible result, and it borrows the codebook from the other side of the argument. The codebook is the twelve functional layers of **The Agentic State** vision paper (Ilves, Kilian, Parazzoli, Peixoto, Velsberg; launched at the Tallinn Digital Summit, 2025-10-09 — see `costs-sources.md` §9). Using an opponent's framework rather than one written here removes the obvious objection that the categories were built to produce the finding. ## Variables ### `infra_ref` — 0/1 · **mechanical, headline-bearing** Does the design refer, in visible text or image, to **any** technical capacity of the state: software, a model, data systems, compute, networks, surveillance cameras, drones, automation? This is a presence test on transcribed content — the same family as `antagonist`, which measured α = 0.82, the most reliable variable in the v1.0 rubric. Published claims should rest on this variable and no other. Coding rule: code `1` for any named technical artefact or system, however mundane, including security hardware. Code `0` for metaphorical or purely institutional language ("bygga Sverige starkt", "infrastruktur" meaning roads and rail). ### `as_layer` — layers `1`–`12`, `;`-separated, or `none` · **interpretive, illustrative only** Through which layer(s) of the Agentic State framework would the state have to deliver this poster's promise? Assigned using the paper's own worked examples — a promise about healthcare waiting times is coded L1/L2 because the paper's Italy case study is agents managing CT-scan waiting lists. **This variable is single-coder and has not been double-coded.** Expect disagreement comparable to `stance` (α = 0.51), and for the same reason: reasonable coders draw the boundary between "a political good" and "an administrative delivery mechanism" in different places. Quote the aggregate shape, never a single design's layer. Layers: 1 Public Service Design/UX · 2 Government Workflows · 3 Policy- and Rule-Making · 4 Regulatory Compliance and Supervision · 5 Crisis Response · 6 Public Procurement · **7 Agent Governance · 8 Data and Privacy · 9 Tech Stack · 10 Cyber Security and Resilience · 11 Public Finance · 12 People, Culture and Leadership**. Layers 1–6 are the paper's *implementation* layers; 7–12 its *enablement* layers. ## Result - `infra_ref = 1` for **2 of 80** designs: M-WC-06 (*Fördubblat kameramål*) and M-WC-08 (*Ett svenskt drönarvärn*). Both are security-register hardware. - **0 of 80** name a model, a datacentre, a cloud contract, compute, or the ownership of any of them. - Layer coverage: **47 design-assignments to implementation layers (1–6), 4 to enablement layers (7–12)**. L7 Agent Governance, L9 Tech Stack and L12 People and Culture are empty. The paper's own rule — *"any deployment of agents within an application layer requires careful alignment across all enablement layers"* — is what gives that asymmetry its force: the campaigns promise outputs from layers 1–6 while saying nothing about the layers that would have to carry them. ## What would make this citable Same standard as v1.0: a second, blind coder on a ~30-design subsample, with α reported per variable. `infra_ref` should score very high (it is nearly mechanical); `as_layer` should be expected to score poorly and be presented accordingly. --- # v3 addendum — machinery coding for long-form platform documents Added 2026-09-05. Extends the v2 machinery coding from posters to the eight 2026 *valmanifest* archived in `valmanifest/` (provenance: `valmanifest-sources.md`). Applied by `build_manifest_dataset.py`; output in `manifest-coded.csv` and `manifest-aggregations.txt`. Does not alter v1.0 or v2. ## Why the unit of analysis has to change A poster is one utterance, so one design = one unit. A manifesto is 2 000–27 500 words, so "one document = one unit" throws away everything and "one proposal = one unit" is neither well-defined across eight differently-structured documents nor reproducible. v3 therefore uses **two units at two levels of trust**: - **The document**, for frequency. Mechanical, reproducible by re-running the script. - **The passage**, for framing. A passage is a contiguous run of machinery language: consecutive lexicon hits less than **250 characters** apart are merged into one. That threshold is a rubric parameter, chosen because the passage count is *stable* there — 45 passages at a 90-character gap, 32 at 250, 31 at 400. It sits on the plateau, not on a slope. Corpus total: **39 passages.** ## Variables ### Stage 1 — mechanical - `hits` — lexicon matches per document. - `hits_per_1k` — **the comparable figure.** The corpus varies fourteenfold in length (C 27 541 words, V 1 955); raw counts are a length measurement before they are a political one. Never quote a raw cross-party count. ### Stage 2 — interpretive, single-coder, **not** double-coded - `frame` — how the passage frames the machinery. Twelve values, grounded in the corpus and listed in the script: `competitiveness`, `capacity`, `dependency`, `labour`, `admin-efficiency`, `regulation`, `rights`, `risk-social`, `risk-security`, `risk-energy`, `inclusion`, `culture`, plus `toc` for a table-of-contents artefact (excluded). - `as_layer` — the Agentic State's twelve layers, exactly as in v2. - `own_claim` — **0/1: does the passage assert public ownership or control of models, weights, compute or data infrastructure?** This is the variable the whole exercise exists to test, and it is the one to state plainly. Expect `stance`-grade disagreement on `frame` and `as_layer` (v1.0 measured α = 0.51 for `stance`). Quote the shape, not a single passage's code. ## What the first lexicon missed — read this before trusting any lexicon result The first pass used a **core** lexicon only: unambiguous machinery words (`AI`, `algoritm`, `språkmodell`, `datacenter`, `beräkningskraft`, `molntjänst`, `öppen källkod`, `öppna vikter`, `automatiser*`, `digitaliser*`). It deliberately excluded high-frequency ambiguous words, and an audit of those exclusions confirmed the exclusions were right: `data` is mostly waiting-list statistics, `digital` is mostly *digital vård*, `teknik` is mostly military or educational, `plattform` collides with *valplattform*, `upphandling` is mostly procurement of welfare services. **But the audit also found a false negative, and it was the most important passage in the corpus.** Moderaterna write: > "Minska Sveriges beroenden av utländsk teknik i kritiska system genom stärkta incitament > och strategisk upphandling för svensk och europeisk teknik inom **säkerhetskänslig IT, > moln och drift.**" The core lexicon missed it because M write `moln`, not `molntjänst`. A second lexicon tier (`INFRA`) was therefore added for infrastructure- and dependency-language: `moln`, `halvledare`, `beroende* av utländsk`, `utländsk teknik`, `säkerhetskänslig*`, `kritiska system`, `digital infrastruktur`, `datadelning`, and the `cyber*` family. That raised the corpus from 32 to 39 passages and changed the substantive finding. **The lesson is the general one: a presence test is only as good as its lexicon, and its lexicon is a judgement call wearing mechanical clothing.** Any future extension of this rubric must audit its own exclusions before publishing a zero. ## Results ### Stage 1 | party | words | hits | passages | per 1 000 words | |---|---|---|---|---| | S | 7 087 | 14 | 9 | **1.98** | | C | 27 541 | 48 | 16 | 1.74 | | L | 9 478 | 12 | 7 | 1.27 | | M | 16 811 | 14 | 6 | 0.83 | | KD | 4 646 | 1 | 1 | 0.22 | | **SD** | 3 057 | **0** | **0** | **0.00** | | **V** | 1 955 | **0** | **0** | **0.00** | | **MP** | 3 222 | **0** | **0** | **0.00** | | total | 73 797 | 89 | 39 | 1.21 | **SD, V and MP contain no machinery language at all** — and that holds under the extended lexicon, so it is not an artefact of word choice. Note that normalising reverses the raw ranking: S is the densest, not C. Absent from all eight manifestos, in ~73 800 words: **`språkmodell`, `öppen källkod`, `öppna vikter`, `leverantörsberoende`, `digital suveränitet`.** ### Stage 2 Frames: `labour` 7, `risk-social` 7, `competitiveness` 5, `risk-security` 5, `admin-efficiency` 4, `capacity` 3, `dependency` 2, and one each of `regulation`, `rights`, `risk-energy`, `inclusion`, `culture`. Layers: implementation (1–6) **13**, enablement (7–12) **18**, no layer 18. L10 Cyber Security is the most-touched layer (7), then L9 Tech Stack (5). **L11 Public Finance is empty.** **`own_claim` = 0 of 38 coded passages, across eight manifestos and 73 797 words.** ## What this does and does not license - **Licensed:** "Across the eight 2026 manifestos — 73 797 words — no passage claims public ownership or control of models, weights, compute or data infrastructure." Mechanical variable plus an explicit, auditable definition. - **Licensed:** "SD, V and MP contain no machinery language at all." - **Licensed:** the absent-terms list. - **Not licensed:** treating this as confirmation of the poster finding. It is a *correction* of it — see below. - **Not licensed:** any single passage's `frame` or `as_layer` as fact. ## How this changes the poster finding The poster coding (v2) returned **47 implementation-layer assignments against 4 enablement**, and that asymmetry carried the argument: the campaigns promise outputs from layers 1–6 while saying nothing about the layers that would have to carry them. **At manifesto length that asymmetry disappears — 13 implementation against 18 enablement.** The parties are not ignorant of the enabling layers. M proposes a national cyber shield for municipalities and strategic procurement to reduce dependence on foreign technology in security-sensitive IT, cloud and operations. L wants Europe to build its own capacity in semiconductors, cloud and AI. C proposes a national AI centre gathering research, compute and industry. These are real infrastructure politics, and the poster corpus could not see them because a poster has no room for them. So the honest finding is narrower and harder than "they never mention the machinery": > They mention it. They propose to buy it, to defend it, to attract it and to be skilled at > using it. Across 73 797 words, not one passage proposes to **own** it. The two passages that name technological dependency as a problem — M-04 and L-04 — both sit in a **defence** chapter, and their instruments are **procurement** and **EU capacity**. That is the Agentic State pattern exactly (`costs-sources.md` §9): the constitutional question, filed under upphandling. ## What would make this citable Same standard as v1.0 and v2: a second, blind coder over the 39 passages, α reported per variable. `own_claim` should score very high — it has a sharp definition — and `frame` should be expected to score poorly and be reported as such.