# Probability adapter evidence The AI and MTTH Scenario Analyzer targets Hearts of Iron IV **Operation Postern 1.19.2.0 (checksum d245)**. Adapter behavior is versioned as `hoi4-1.19.2.v1`; results always include that adapter version and game build. This review used the installed game's documentation, vanilla source examples, and an offline Paradox wiki snapshot. Paths below are relative to a Hearts of Iron IV installation or to the offline snapshot. SHA-256 values identify the exact files reviewed on 2026-07-22. ## Adapter semantics | Adapter | Proven interpretation | Exact probability conditions | Deliberate boundary | | ------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------------------------------------------------------------------------------------------------------------------- | | `event_mean_time_to_happen` | Ordered `base`, `add`, and `factor` operations produce an effective duration. When the trigger is active, `M` days is the median of daily checks, so daily hazard is `1 - 2^(-1/M)`. Inactive event triggers are polled every 20 days. | A time-horizon chance is exact only while trigger state, every modifier input, scheduled change, and any inactive-to-active polling phase are declared. | It does not execute events or infer state changes. Missing poll phase produces a 0–20 day activation bound. | | `event_option_ai_chance` | Options form one proportional categorical pool. An omitted block has weight 1. An all-zero pool selects the first option. | Every option and eligibility result for the event must be present. | The engine's d100 selection granularity is reported for effective shares below one percent. | | `decision_ai_will_do` and `mission_ai_will_do` | Ordered source modifiers produce a raw desirability score after eligibility. | No normalized probability is claimed. | Local documentation does not prove a complete categorical denominator or a universal selection formula. | | `national_focus_ai_will_do` | Each eligible focus receives an independent uniform draw from zero to its evaluated score; the maximum wins. A just-completed prerequisite and AI strategy focus factors can change the score. | The complete eligible candidate pool and every external focus factor must be declared. Exact probabilities use the independent-uniform maximum race, not `weight / sum(weight)`. | Ordered AI strategy plans can override weighted choice and are reported rather than simulated. | | `technology_ai_will_do` and `doctrine_ai_will_do` | Eligible candidates use the independent-uniform maximum-score race. Research factors, cost, date, bonuses, and forced strategy choices can alter or override ranking. | The complete eligible pool and all external research factors must be declared. | Research duration is not selection probability. Forced research strategies and wider research-slot scheduling are not simulated. | | `direct_random` | `chance` is an independent percentage in `[0,100]`. | The chance expression must resolve. | It is never normalized against neighboring blocks. | | `random_list` | Entries form the complete local proportional weight pool. Nested entries retain their immediate conditional share and multiply through enclosing lists for a full path probability. | Every entry, enclosing list, and modifier on the selected path must resolve. | Dynamic entry identifiers, dynamic parent paths, or effect-derived values remain unresolved unless supplied explicitly. | | `ai_strategy_factor` | Supported strategy types modify a target score. `research_weight_factor` is a percentage factor; positive `research_tech` forces prioritization. | Raw factor evaluation is exact when the strategy and target are explicit. | A strategy definition is not itself normalized into a probability pool. | | `custom_weighted_pool` | The manifest declares categorical or independent selection plus recovery, caps, cooldowns, removals, resets, timer changes, and terminal states. | Exact finite-state analysis is used while the state space stays bounded; otherwise the result is explicitly beam-bounded or seeded Monte Carlo. | Only declared state transitions run. Gameplay effects and campaign state are never inferred or executed. | ## Installed documentation reviewed | Relative path | SHA-256 | Used for | | ----------------------------------------------- | ------------------------------------------------------------------ | ------------------------------------------------------------- | | `launcher-settings.json` | `f99265973bba12f30906cc98629d8520218ff8cd90a5f9d1d82e7f00fb737b9e` | Installed game version and checksum | | `common/decisions/_documentation.md` | `5e7e686b5582f32c38d898d805e17171e88f3bfeee4a030deb1884f9726de019` | Decision scopes, eligibility surfaces, and evaluation cadence | | `common/doctrines/_documentation.md` | `f0d1b7ad1f1dc8afe59942a406784a2d72c0c73dd42963c2502f89961eea85d4` | Current doctrine definitions and AI fields | | `common/script_constants/documentation.md` | `e2812efbddc7961038e1f1f2b6dd6b059bdc70704ec87d1fe530835d9141be1c` | `constant:` schema and nested constant paths | | `documentation/script_concept_documentation.md` | `884bfe9f9207ea4b25b36b5129d0f77d11d4b2f7eb0592587767ec58d24930c7` | Script constants and dynamic script concepts | | `documentation/effects_documentation.md` | `2cdc93d26775990468eed167442f7a416b13059fb8fe07d139b54250b79209af` | `random_list` and AI strategy effect syntax | | `common/ai_strategy/_documentation.md` | `70f5d755ae7ed4060e49cc64a393d1a0dc5611b3acc8abc6dfb4538781c915a1` | `research_tech` and `research_weight_factor` behavior | The offline wiki pages `AI modding`, `Data structures`, `Decision modding`, `Event modding`, `National focus modding`, `Scopes`, and `Technology modding` were reviewed in parallel. Their respective SHA-256 values were `f1fdb9fd…a14f`, `20a03385…a14f`, `500e8161…68c`, `b0186d9b…e00`, `b1c7f624…6faf`, `04866815…fdc`, and `48050261…fdc`. ## Vanilla examples reviewed | Surface | Relative path and line | SHA-256 | | ----------------------------- | -------------------------------------------------------------- | ------------------------------------------------------------------ | | Event MTTH | `events/Yugoslavia.txt:49` | `138279b294ef0d567e477efe2f0ebb72c5c0a7f2920ed3fd03d2792749624dbc` | | Event option `ai_chance` | `events/AAT_Generic_Events.txt:44` | `e718d19edba2d56c75a5bbf1fa0c51bbb6cda54e0919434c4490fec9dabcea69` | | Decision `ai_will_do` | `common/decisions/_generic_decisions.txt:56` | `83cd276f8fd2c0c90470225c996d7afee7b8c08437289dd23d819a4a29edb187` | | Focus `ai_will_do` | `common/national_focus/austria.txt:122` | `0f197e321cb9e0c267a11c78bccb570ef42d1fc251e7af05fb4f385de27a1059` | | Technology `ai_will_do` | `common/technologies/industry.txt:30` | `06d45f11a0bfdee35ef72b935b2bee3b0083a49eb771e5c6b9391c087ed7d94f` | | Current doctrine `ai_will_do` | `common/doctrines/grand_doctrines/land_grand_doctrines.txt:14` | `cd81a877366c582f096422320cd7fcf9fb259afcc8d45d3dbb887b37792b329d` | | `random_list` | `events/AAT_Sweden.txt:1760` | `8fad5c10d147afab27eb42c78f3261c804cd6149f1c81927418ab0bbe3b8bf29` | | Direct `random` | `events/BBA_Switzerland.txt:1197` | `4331bbe2ea7d0844eb357ecd73f915c1ba3e5a0b115ae62ea5b95a8011e1145e` | | AI research strategy factor | `common/ai_strategy/ENG.txt:3206` | `f6fed5347c68ab2a3b79e4864dc5d72490c9cabe46a16e865932d2e5a37716d9` | The installed vanilla inventory contained 71 event files with MTTH, 101 with `ai_chance`, 86 decision files with `ai_will_do`, 76 national-focus files, 13 technology files, 19 doctrine files, 38 event files with `random_list`, and 9 AI strategy files with `research_weight_factor`. These counts were used to select examples and to test discovery breadth; they are not runtime assumptions. ## Unsupported constructs When a game root is configured, adapter execution verifies `launcher-settings.json` against the supported raw version and checksum and fails closed on a missing, unreadable, or different build. Without a configured game root, results identify the adapter target without claiming that an installed build was verified. The analyzer reports, but does not guess or execute: - arbitrary effects used to calculate later weights; - scope chains, event targets, game rules, DLC state, dynamic identifiers, meta effects, or scripted-localisation results that are neither present in source nor declared through structured scenario bindings; - runtime scope pools whose candidate catalog is not declared; declared catalogs are filtered and enumerated exactly, while normalized shares additionally require explicit completeness; - parameterized scripted triggers whose arguments cannot be resolved from the selected source; - recursive MTTH variables or helper cycles; - incomplete focus, technology, doctrine, event-option, or random-list candidate pools; - categorical or discrete correlation requests; numeric ranges and numeric distributions are supported by the correlation sampler; - invalid or non-positive-semidefinite correlation matrices; - full strategic AI planning, research-slot scheduling, event-effect execution, or campaign simulation.