# 900 — Decisions of record

Approved by Standa in this chat on **2026-10-12**. Preserve these without repeatedly reopening them. If implementation reveals a real conflict, explain it and ask in ordinary final chat.

## Approved product decisions

| ID | Topic | Decision |
|---|---|---|
| D01 | Mode | Add Příběh beside Milionář; retain both |
| D02 | Gameplay | Four A–D options; answer must match the actual selected story |
| D03 | Lives | Three; wrong answer removes one and reveals the real event; third wrong ends the game |
| D04 | Continuity | Continue the same canonical plot after errors; no alternative branches |
| D05 | Knowledge | Generate from ordinary LLM knowledge; no online search initially; accepted accuracy limitations for newer works |
| D06 | Future research | Document online verification/grounding as a later optional phase |
| D07 | Work selection | Exactly the existing Millionaire selection and launch contexts; no new selection system |
| D08 | Difficulty | Easy 10 / Medium 15 / Hard 20; major events to finer plot details; every variant covers the whole story |
| D09 | Questions | Plot only; no actor/production/general trivia |
| D10 | Viewpoint | Often protagonist, but any relevant character/event as the scene requires |
| D11 | Progress | Five spoiler-free narrative chapters replacing the money ladder, scene progress, visible lives |
| D12 | Checkpoints | Visual only; no life restoration or checkpoint saves |
| D13 | Preparation | Generate the whole pack before play: plot flow, questions, answers/reveals and ending |
| D14 | Voice of prose | Continuous concise narrator/book-like retelling; no game presenter |
| D15 | Text fit | Short scene text fitting the current cards; existing autoscroll as fallback |
| D16 | TTS | Reuse existing voice configuration; default test path is text-only, no voice |
| D17 | Name | Czech mode label Příběh; English code name story |
| D18 | Event order | Original audience/presentation order, including flashbacks/revelations |
| D19 | Defeat | Stop at current event, reveal it, do not reveal the rest of the plot |
| D20 | Completion | Brief canonical ending and results with correct answers, reached chapter, remaining lives; no money/bonuses |
| D21 | Launch choice | Selector Milionář / Příběh on the shared existing modern title card |
| D22 | Exclusions | Persons and mixed-title topics remain Millionaire-only |
| D23 | Mode memory | Remember the last selected mode |
| D24 | Lifelines | None in Story: no 50:50, Tip, or Audience |
| D25 | Difficulty preference | Story has its own remembered difficulty; voice shared; no presenter-style settings |
| D26 | New attempts | Every LLM game is freshly generated; no generated-content memory/cache/reuse |
| D27 | Demo repeatability | Authored demos always stay the same |
| D28 | Layout | Closely follow modern Millionaire: narration/question/answers left, story chapters right, lives visible |
| D29 | Evaluation | Immediate result; retain chosen/correct/wrong visual cues; no suspense countdown |
| D30 | Pace | Player presses Continue after reading; no automatic progression or forced wait |
| D31 | Loading | Existing loading style; show first scene only after complete pack preparation/validation |
| D32 | Insufficient knowledge | Return an honest limitation/error, offer Retry/Back; prompt forbids made-up substitute plot |
| D33 | Demo target | Original Matrix (1999), fixed Story variants at all three counts |
| D34 | Engineering | Clean modular code, no spaghetti/duplication; reuse existing infrastructure; consistent established design |
| D35 | Initial planning task | Create the English multi-file plan; subsequent implementation approval is recorded in D37 |
| D36 | Plan maintenance | Include goals, current state, steps, lessons, progress and handoff so another agent can continue |
| D37 | Implementation approval | Implement the whole agreed feature and Matrix demo, then verify on that demo |
| D38 | Test sequence | Authored text-only demo first; Standa tests next; only after that run Luna versus Sonnet and tune reasoning/prompts |
| D39 | Resume — 2026-10-13 | After the requested full deploy and pause, Standa explicitly asked to resume and finish the feature autonomously |
| D40 | Additional demos — 2026-10-13 | Finish the complete functional game/integration first; finish further demos afterwards, not as a prerequisite or parallel scope expansion |
| D41 | Demo acceptance / paid testing — 2026-10-13 | Standa confirmed he tested the game and explicitly authorized real Luna/Sonnet testing and optimization; the D38 gate is satisfied |
| D42 | Complete demo catalog — 2026-10-13 | Standa requested a Story counterpart for every existing Millionaire demo after implementation and AI optimization were completed; the D40 deferral is satisfied |
| D43 | Series scope — 2026-10-13 | Standa approved the recommendation to use complete closed narratives versus clearly identified episodes for episodic series; exact default selections are recorded in I23 and `140` |
| D44 | Harry Potter collection — 2026-10-13 | Standa requested every movie from the Harry Potter collection as demos in both Millionaire and Story; add the seven missing films and preserve the existing first movie |

## Implementation defaults and derived rules

These fill routine technical details; they were not separate user votes. They may be adjusted with evidence if product behavior remains identical. Record substantive changes in the handoff rather than silently rewriting approved decisions.

| ID | Detail | Default / reason |
|---|---|---|
| I01 | Plan folder | `docs/plan/story-game/`, English filenames/prose as requested |
| I02 | Preference IDs | `game_last_mode`, `story_difficulty`; check collisions before implementation |
| I03 | First-use preference | Millionaire mode, Medium Story difficulty; invalid values use those defaults |
| I04 | Ineligible stored mode | Use Millionaire for that launch without overwriting the stored Story choice |
| I05 | Chapters | Five fixed local labels; equal groups of 2/3/4 scenes derive from 10/15/20 |
| I06 | Score interpretation | Correct and answered counts distinct; wrong answers still traverse the plot |
| I07 | Final precedence | Zero lives loses even on the last question; otherwise reaching the last scene wins with mistakes allowed |
| I08 | AI call budget | One inference per explicit Start/replay/retry; no automatic repair or feedback inference |
| I09 | Failure cost | No life loss for preparation/validation/provider errors or cancelled work |
| I10 | Source lifetime | Only active pack in memory; no story resume/history; usage metadata/preferences and authored fixtures are separate |
| I11 | Demo eligibility | All 29 catalog IDs now have both authored sources; require all three local Story packs; unknown/incomplete IDs never cause paid fallback |
| I12 | Ongoing series | Cover the known selected endpoint without inventing a future finale |
| I13 | Existing season context | Preserve if supplied; no new season picker and no guessed scope |
| I14 | Legacy/classic | Canonical shared launch explicitly uses modern for both modes; preserve explicit classic compatibility without deleting/migrating its setting or adding classic Story |
| I15 | Code module map | Implemented small Story package in `030`; adapt later refactors without changing ownership |
| I16 | Field limits/geometry | Initial values in `050`/`070`, verified and adjusted together using real text/layout evidence |
| I17 | Production language | Czech generated narration like current prompts; interface cs/sk/en; docs/code names English |
| I18 | Intro | No extra paid greeting/presenter inference; title/rules -> loading -> first narrated scene |
| I19 | Verification | Existing stubs/probe/harness reused; focus on fixed Matrix text checks and affected Millionaire regressions |
| I20 | Review diagnostics | Capture only the explicitly requested Matrix benchmark's prompt/response/pack under the existing harness output folder for manual plot review; no credentials/headers, gameplay cache or production archive |
| I21 | High-effort room | Story-only 32K output / 300 s for OpenAI/Anthropic high/xhigh/max, justified by observed truncation/timeouts; 16K/client defaults otherwise; no automatic repair or global setting change |
| I22 | Measured recommendation | Sol 6.1 / medium for the Story trial after fresh 10/15/20 confirmation; recommendation rather than a changed product default; retain user-selected provider/model/effort and documented accuracy limits |
| I23 | Authored series scopes | Complete Breaking Bad, Chernobyl and Game of Thrones; identified aired S01E01 for Friends, House, Sherlock, Stranger Things, Big Bang Theory, Mandalorian, Office US and Simpsons. These individual selections implement the approved mixed-scope recommendation; they were not separate user votes |
| I24 | Scope ownership | One Story-only `scopes.json` overlay, preserved explicit caller scope, original Millionaire metadata and existing source-note label; no copied catalog/picker or runtime fixture generator |
| I25 | Film collection implementation | The original eight Harry Potter movies (TMDb collection 1241), separate Deathly Hallows Parts 1/2, IDs `harry-potter-1`–`8`; reuse existing stable catalog order to group them without altering old data, engine, settings or dialogs |

## Authorization record

- Read-only investigation of Millionaire and discussion: authorized.
- Creation of this plan: explicitly authorized (“do it”, English, multiple files, takeover/lessons included).
- Implementation, actual game/API/TTS runs, GUI/device tests, commits and deploys: not performed under this planning task. Verify the next actual chat instruction before beginning new work.

## Future changes

Append date, exact user instruction or evidence, superseded decision ID, new rule, and affected chapters. Do not treat an agent's convenient workaround as a product decision or report a requested plan as an implemented feature.

### Subsequent implementation authorization

The active goal authorized implementation and Matrix demo testing. Standa confirmed his successful hands-on test and authorized real Luna/Sonnet optimization in D41. Earlier planning-only and blocked notes are historical, not the current scope.

After that work completed, Standa requested all demo counterparts (D42) and answered “Ano” to the proposed complete-narrative/identified-episode policy (D43). Authored catalog completion and targeted offline verification are authorized. No new paid calls, GUI/device tests, commit or deploy are implied by this request.