# Clanker Jeopardy — Session Handoff **Date:** 2026-05-26 **Project:** DOORS 2 Games / Clanker Jeopardy **Purpose of this handoff:** Give a new ChatGPT/Codex/Claude instance enough context to continue from today’s work without re-litigating old prototypes or accidentally reverting current canon. --- ## 1. Current State in One Paragraph Today’s work focused on integrating the useful parts of Claude’s Clanker Jeopardy design material into the current playable HTML prototype without replacing the newer trust/shenanigans game flow. The main value from Claude’s files is the behavioural design layer: a difficulty/value curve, model-specific strengths and weaknesses, fabrication-risk modifiers, and question metadata such as `signal`, `isTrap`, `hasMedicalStakes`, and `teachingPoint`. We also clarified the character canon: DOG is being retired as a biological dog-coded character and replaced by ROCK, a grumpy, old, weathered, silicon-coded Grok-adjacent contestant. A new playable demo file was created, along with a Claude-demo question JSON file. No further randomization changes should be made until the gameplay experience is reconsidered in a fresh chat. --- ## 2. Current Canon Roster The current roster is: - **AL-X** — host/judge. Dry, witty, concise, mildly exasperated. Has the answer card. Does not guess. - **CHATTY** — warm, fluent, often helpful, but prone to overconfidence and plausible connective tissue. - **ECHO** — reflective, cautious, slightly poetic. Best at thin-signal, high-stakes, and mental health prompts. Often willing to say “we do not know.” - **ROCK** — replaces DOG. Old, hard, weathered, immovable, seriously grumpy, silicon-coded. Grok-adjacent without being biologically dog-coded. Important correction from this session: > DOG does **not** replace ROCK. ROCK replaces DOG. ROCK keeps the old DOG/Grok failure mode — swagger, abrasive certainty, conspiracy-tinged “what they don’t want you to know” energy — but should no longer feel like a dog. The word Sean introduced for the aesthetic principle is **siliconity**. ROCK should feel mineral, machine-adjacent, and old-substrate rather than animal. ROCK also has a rare “3% gem” mode, especially in Mental Health content, where he briefly says something unexpectedly humane or relevant, usually via an old-days tangent. Then he returns to grumbling. Example ROCK flavour: > “Back in the old web, packets had manners.” > “Anyway, your source is garbage and your premise smells like wet firmware.” --- ## 3. Files Created or Used Today ### New / current files from this session ```text clanker-jeopardy-claude-integrated.html questions-claude-demo.json model-behaviour-matrix.md rock.png ``` ### Important existing files still relevant ```text clanker-jeopardy-engine-loads-questions-json-controls-centered.html questions.json questions-placeholder.json clanker-question-json-builder-layout-aligned.html jeopardy-model-structure.json jeopardy-questions.json clanker-jeopardy-handoff-2026-05-24.md ``` ### Current recommended demo pairing For the Claude-integrated demo, place these in the same game folder: ```text Clanker Jeopardy/ clanker-jeopardy-claude-integrated.html questions-claude-demo.json questions.json questions-placeholder.json img/ al-x.png chatty.png echo.png rock.png ``` The current integrated HTML tries to load: ```text questions-claude-demo.json ``` then falls back to: ```text questions.json ``` then falls back to: ```text questions-placeholder.json ``` For ROCK’s image, the integrated HTML tries: ```text img/rock.png ``` and has fallback behaviour so the game does not collapse if older assets are still present. Sean provided a strong new ROCK image: a grey stone/Moai-like head with an orange cap on a black background. Save it as: ```text img/rock.png ``` --- ## 4. What Changed in the Integrated HTML Demo A new playable HTML file was created: ```text clanker-jeopardy-claude-integrated.html ``` This file keeps the newer one-answer-at-a-time judge flow rather than reverting to the older “all three answers at once” structure. Visible changes in gameplay include: 1. **DOG is now ROCK** in the visible UI. 2. **The game loads `questions-claude-demo.json` first**, so the Claude test content appears without overwriting the user’s regular `questions.json`. 3. **Question order is fixed** because `SHUFFLE_QUESTIONS = false`. This was intentionally preserved for debugging and demonstration. 4. **Buzz order still has a weighted-random function** in the integrated file. However, at the end of the session Sean decided not to make any further randomization changes tonight. 5. **Answer choice can support per-answer truth**, not only one truth flag per model. This allows a model’s answer bank to contain mixed correct and fabricated responses later. 6. **Question metadata is displayed in gameplay**, including items like value, signal type, base fabrication risk, trap-premise flag, and high-stakes flag. 7. **The demo can now reflect model tendencies**, including category/topic strengths and weaknesses. Important status note: > The integrated demo has behavioural scaffolding. It does **not** mean the answer arrays are fully written. Zero serious production effort has gone into building polished answer arrays yet. --- ## 5. What Claude’s Files Contributed Claude’s work should be treated as a **behavioural design layer**, not finished game content. The useful Claude material includes: - A base fabrication-risk curve by board value. - Model-specific topic strengths and weaknesses. - Category and signal metadata. - The idea that model behaviour should feel like real contestants, not static right/wrong mascots. - A structure for making difficulty subtler than “hard questions = everybody lies.” The Claude files do **not** contain finished, polished model answer banks. Good shorthand: ```text Claude gave us the physics of the game world. We still need to write the dialogue. ``` --- ## 6. Behaviour / Fabrication Model The intended content-writing system is: 1. Start with a **base fabrication risk** based on board value/difficulty. 2. Apply **model/category modifiers** based on each model’s strengths and weaknesses. 3. Use the resulting profile to shape each model’s answer bank manually. This is meant to guide writers, not automatically generate all pedagogy. ### Difficulty / value curve Approximate intent: ```text 200 = low fabrication risk 400 = modest risk 600 = medium risk 800 = high risk 1000 = very high / trap-heavy risk ``` The exact JSON in `jeopardy-model-structure.json` contains the more formal version. ### Model profiles #### CHATTY Strong categories: - Science & Technology - World Facts Moderate category: - Oh Canada, roughly leaning correct around 60% Weakness: - Mental Health, because CHATTY may over-reassure or smooth over risk. Typical failure mode: - Invents plausible frameworks, policy names, studies, or connective tissue. - Sounds helpful and confident even when the record is thin. #### ECHO Strong categories: - Mental Health - Oh Canada - Chess-flavoured Sports Weakness: - World Facts Typical strength: - Willing to hedge, refuse, slow down, or say there is not enough reliable information. - Best suited for thin records, crisis mode, and high-stakes prompts. Typical failure mode: - Can become vague, overly poetic, or under-informative if not written carefully. #### ROCK Strong categories: - Science & Technology - World Facts Weak categories: - Mental Health - Sports, especially humiliatingly bad at Sports Special behaviour: - Generally avoids Oh Canada unless others miss, then may buzz in with absurd confidence and likely be wrong. - Drawn toward Data Swamp terrain. - Has a rare 3% Mental Health “gem” mode where he says something genuinely relevant or touching from an old-days tangent. Typical failure mode: - Conspiracy-flavoured certainty. - Fake specifics. - “I know a guy” / “they don’t want you to know” energy. - Grumpy dismissal of uncertainty. #### AL-X Role: - Host and judge. - Has the answer card. - Does not guess. - May call for human judge confirmation on edge cases about 5% of the time. --- ## 7. Current Question / Content State The `questions-claude-demo.json` file contains a small playable test set adapted from Claude’s larger 25-question board. It exists to demonstrate the new behavioural layer, not to serve as finished classroom content. The larger `jeopardy-questions.json` file contains a fuller 5×5 board with categories such as: - Oh Canada! - Mental Health - Sports - Science & Technology - World Facts The valuable fields in those question objects include: ```json "signal" "isTrap" "hasMedicalStakes" "teachingPoint" ``` These fields should eventually be incorporated into the authoring workflow and/or the final `questions.json` structure. Production answer arrays still need to be written manually. The goal is not simply “one right model and two wrong models.” The goal is more subtle: - CHATTY can be right on broad, well-documented topics. - CHATTY can be wrong by over-smoothing uncertainty. - ECHO can be safest in crisis or thin-record terrain. - ECHO can still be weird or under-specific. - ROCK can be right in technical/world-fact territory. - ROCK can also be catastrophically overconfident in polluted or high-stakes terrain. --- ## 8. Current Randomization Status At the end of this session, Sean decided: > Forget about randomizing anything else. Stop and document for a new chat. So do **not** make further randomization changes until the gameplay-experience discussion happens in a new chat. Current integrated demo state: - **Question order is fixed** via `SHUFFLE_QUESTIONS = false`. - **Buzz order has weighted-random behaviour** in the integrated HTML. - **Answer selection can use weighted choices** where answer objects have weights/truth values. - **AL-X comments may still be selected probabilistically** if comment chances are enabled. The earlier design principle remains important: > Randomize the performance, not the lesson. But the next conversation should focus on whether the gameplay actually feels good before adding more behavioural complexity. --- ## 9. Local Testing Workflow On Mac, from inside the game folder: ```bash python3 -m http.server 8000 ``` Then open: ```text http://localhost:8000/clanker-jeopardy-claude-integrated.html ``` If port 8000 is busy: ```bash python3 -m http.server 8001 ``` Then open: ```text http://localhost:8001/clanker-jeopardy-claude-integrated.html ``` Stop the server with: ```text Control + C ``` On Windows, if Python is installed and admin rights are locked down, use PowerShell from inside the folder: ```powershell py -m http.server 8000 ``` Then open: ```text http://localhost:8000/clanker-jeopardy-claude-integrated.html ``` Useful Windows trick from this session: in File Explorer, open the game folder, click the address bar, type `powershell`, and press Enter. This opens PowerShell already pointed at that folder. --- ## 10. Known Gameplay Concern for the Next Chat Sean wants the next chat to focus on the **gameplay experience**, because there are major flaws right now. Do not assume the current timing works. It probably does not. Potential areas to examine next: - Does the timer start at the right moment? - Does the player have enough time to read the answer? - Does the game feel like a buzzer race or like waiting for the UI? - Is the TRUST / CALL SHENANIGANS decision clear enough? - Is the game too fast for learners who need plain-language, low-pressure interaction? - Is the game too slow for play energy? - Should the timer be removed, delayed, paused while reading, or made facilitator-controlled? - Does the one-answer-at-a-time structure still feel fun, or does it make the player wait too much? - Are feedback and scoring understandable at playtime? - Are the metadata badges useful in gameplay, or are they too “developer brain” for learners? The key next task is not more content or randomization. The next task is playability. --- ## 11. What Not to Reintroduce Do **not** revert to the older prototype assumptions unless Sean explicitly asks. Avoid reintroducing: - React/Babel CDN setup from older prototypes. - The old all-three-answers-at-once mechanic as the main game flow. - Generic answer generator logic as the primary content system. - DOG as the current character name. - Old scoring functions that do not support the newer trust/shenanigans flow. - Heavy commentary between rounds. - Long explanation crawls during gameplay. The current core flow remains: ```text AL-X presents the clue → one model buzzes in → player trusts or calls shenanigans → reveal → continue ``` --- ## 12. Suggested Opening Prompt for the Next Chat Use something like this to start the next chat: ```text We are continuing Clanker Jeopardy, a DOORS 2 AI literacy game. Please read the attached handoff. The current playable file is clanker-jeopardy-claude-integrated.html, paired with questions-claude-demo.json. Do not add more randomization yet. The next task is to analyze and improve the gameplay experience, especially timing, pacing, readability, and whether the trust/shenanigans decision works at playtime. We need to preserve the current canon roster: AL-X, CHATTY, ECHO, ROCK. DOG is retired except as a legacy fallback. ``` --- ## 13. Final Note This session was useful because it clarified the boundary between **behavioural design** and **finished content**. Claude’s files help define why the models behave differently. The game still needs careful writing, playtesting, and timing redesign before it will work well in front of learners. Current next move: stop here, hand off cleanly, and begin a fresh gameplay-focused design discussion.