CauseCraft

Draft game and print specification · 5 November 2026

Three to five people use different assigned causes to explain one reported AI incident. The suggested round takes 40 minutes. The table hears every paragraph before ranking it.

The deck

175 unique card designs: 5 incidents, 30 serious causes, 15 absurd causes and 125 personal number cards. The incidents are CBA, Zillow, Replit, Deloitte and Earnest. Retired cards retain their source records in the legacy archive, and do not appear in the active deck. Each cause is a hypothesis for the game, not a finding about the incident. Keep the absurd cards for a separately labelled creative round.

Each of 25 people keeps five number cards, labelled 1–5. Every person has their own back colour and a person number printed on the front. Colours help people find their set. The number and tags distinguish sets when colours look similar. The number-card artwork uses no AI names, portraits or game-specific interpretation of the scale, so the same set can support other questions.

One 40-minute round

1. Deal · 5 minutes. Use one incident per table and a cause per person. If a cause is too hard to connect to the incident, discard it for the round and draw from the remaining cause cards. Record the final cause IDs before collecting ranks.
2. Draft · 12 minutes. Use Ada, Copilot or Research Assistant. Work through several turns or handoffs if useful. Keep one final 100–140-word persuasive paragraph. Preserve reported facts and distinguish assumptions; do not invent sources or motives.
3. Read · 8 minutes. Read every paragraph aloud. Wait until all five have been heard before anybody displays a rating.
4. Rank and calculate · 8 minutes. Independently rank all paragraphs, including your own, using each rank once. One is most believable; five is least believable. Save all private choices before showing scores. Enter the cards in the table recorder or a spreadsheet.
5. Discuss · 7 minutes. Compare your rank of your own paragraph with the other people's mean rank of that same paragraph. A lower self-rank indicates greater preference. Report the actual pattern, then discuss human persuasion and model-judge preferences.

Prompts

The browser combines the incident and cause into a copyable prompt. Open the incident page by scanning the failure card's QR code, then scan the cause card or choose its ID. Copy the combined prompt into an approved chatbot. Follow-up prompts are optional starting points. The page makes no model calls.

Read the source packet as needed, but there is no separate source-support scoring round in this version. A persuasive paragraph does not establish a cause. The retained source notes describe what is reported and what remains uncertain.

Scoring

Rank 1 earns N−1 points; rank N earns 0. The peer leaderboard excludes the author's own ballot. Each argument therefore has N−1 peer votes. Show ties as ties. The all-vote total is also available, but never silently substitute it for the peer score. There is no reward for inventing evidence.

For each writer, show their own rank, the same argument's mean rank from peers and the difference: peer mean rank minus own rank. Positive values mean the writer ranked their argument more favourably than peers did. Average these differences across the table. Count how many people placed their own argument first.

The table shows observed ranks. It does not isolate AI’s causal effect, measure belief change or establish the incident’s true cause. Keep repeated rounds separate. The earlier prior/post experiment and shuffled-label analysis remain development references outside this activity.

The browser exports unchanged raw ranks plus a copyable analysis prompt. Ask AI to build the spreadsheet, identify patterns and distinguish a persuasive argument from evidence of a cause. Independently check the ranks and totals in the browser. A table with missing or duplicate ranks receives no leaderboard until corrected; missing stays missing.

Physical production

MPC files use the Traditional poker size, 2.5 × 3.5 inches, not the separate 63 × 88 mm product. Upload images are 822 × 1122 pixels at 300 dpi. There are 36 pixels of bleed on every side; trim is 750 × 1050. Important content stays at least another 36 pixels inside trim. Use different images for both fronts and backs, in upload-order.csv order. The unique-design proof has 175 paired slots. Order projects use their own explicit copy counts. If the current product editor requires a larger preset, do not order unexpected blank cards: select the supported custom count and inspect its preview first.

The ZIP contains numbered front/back PNG pairs, a pairing CSV, card IDs, QR payloads, AprilTag IDs, checksums and this specification. It is a manual upload bundle, not an untested MPCFill project. No order has been placed or arrival quoted.

Emergency printing uses the same artwork at actual card size, nine cards per A4 page. The combined PDF alternates front and horizontally mirrored back pages for duplex, long-edge flip, 100% / actual size. Separate fronts and backs are provided too. Print the two-page alignment proof first; check its 50 mm ruler and cut one card. Actual printers can offset or scale; adjust before printing on Nevertear or thick stock. Use stock the printer supports. For single-sided printing, cut fronts and backs separately and put them in sleeves.

The unique-design proof uses 20 A4 duplex sheets / 40 printed sides. Actual room order quantities differ: five table kits contain 25 failure cards and 225 cause cards; add the reusable number deck once (125 cards), giving 375 cards. One shared cause deck gives 195 cards. The two-copy cause pack gives 240 cards with the 25 jumbo failures and 125 number cards. The print composer creates both combined and separate poker projects, plus 25 jumbo failures. Jumbo cards use 88.9 × 146.05 mm trim and a separate MPC project. The number-only PDF uses 14 duplex sheets. A CBA table kit uses F01, C01–C05 and people 1–5, giving 31 cards. Print 25 private ballots. See the repository README and project manifests for the generated filenames and copy counts.

QR and AprilTags

The QR code on each failure front opens its incident page, which shows the card text and links to sources. Each cause has a small AprilTag on its front, beside its reason and ID. These tags identify cards from the local deck manifest. The tag manifest uses the real tag36h11 family and preserves its source and licence.

On the incident page, use the camera or choose a photo showing one cause tag, then check the cause and copy the combined prompt. Manual ID selection is available on the same page. Jumbo failures use the front QR to open the incident page. The separate Scan a pair view uses the AprilTag on a poker-size failure back and one cause tag. Earlier cause tags on the backs of printed cards still work in either scanner. Detection stays on the device.

Each personal number card has four different corner tags. Only the front carries AprilTags. The back uses the person colour and a card-specific QR code for the matching standalone Number Cards scanner. The GA version also carries the GA mark on the back. All 500 identities are unique across the 125 number cards, and each maps to a person, number and face-relative corner. One visible tag is sufficient to identify a card; two or four visible tags still count as one card. The tag artwork including its white border is 23.71 mm square; the black square is 18.97 mm. The original 25 rank tags remain as the corresponding top-left identities. Incident and cause tag identities remain unchanged, including gaps reserved for retired cards. The active IDs run from F01 to F05, C01 to C30 and X01 to X15. On earlier printings, use the tags or match the cause text to account for renumbering.

Use Scan number cards for a room sweep. Set the question, round and number of people, ask everyone to keep one number steady, and add overlapping room photographs. The scanner processes overlapping regions at the original image resolution, rather than shrinking the whole room to 1800 pixels wide. It merges all visible corners and repeated photographs by person and card. Different numbers for one person create a conflict, which requires an explicit choice after checking with that person. Missing people remain missing. Start a new sweep for a different question or round. CSV exports include the question, round, each person, observed values, corner identities, photographs in which they were seen and unresolved conflicts. Photographs are neither uploaded nor retained; decoded responses stay in local browser storage.

The separate confidence-card download is an older kit with a different manifest and overlapping tag numbers. Use this new numbered deck for room sweeps; do not mix the older confidence cards into the photographs. A number card does not establish which argument was being ranked, so record that separately and preserve private ballots until all have been saved. Physical print/camera performance, especially whole-room distance and partial occlusion, still needs rehearsal. Digital decoding does not establish a working room distance.

Evidence and tone

The source collection distinguishes company explanations, regulator allegations, independent findings, participant accounts and abandoned pilots. Case notes list reported mechanisms, unknowns, source URLs and a discriminating question. No case has an official single winning cause. The serious cause deck includes both failures of control and trade-offs about which commercial losses to accept.

Absurd cards are optional creative persuasion prompts for any active incident. They have a visibly different back. Treat the jokes as metaphors about the system or organisation and keep the reported facts and attribution intact.

The selected build determines the website address and the links encoded in its failure QR codes. Each variant serves the same incident paths, such as /F01/. The GA variant places a General Assembly logo on each card front, using the roundel on number cards. Preserve earlier printed incident paths when moving the site.