Glossary

hakadoru.ai で使われる主要な用語の定義集です（30件）。

Domain Objects

Item: User-managed knowledge data unit stored in the `items` table. Classified into 3 types: snippet / output / conversation. Character settings, world settings, location settings are all Items.; FragmentKnowledge AssetResource
Fragment: An Item inserted into a prompt as context for the LLM. The term specifically refers to Items in the context of prompt insertion. Users explicitly select which Fragments to include.; Context FragmentKnowledge Asset
Artifact: An Item saved from part or all of an AI response via the Artifact Capture operation. Scored with `deliverable_score` (0-100) and grouped by `deliverable_group`.; Deliverable
Work: A single novel/story unit authored by the user. Contains multiple Scenes. Items belong to a Work scope.; Workspace (different concept)
Workspace: A user-tenant scope. One record per user-brand combination. Created automatically on registration or tenant join. `UNIQUE(user_id, tenant_id)`.; Work (different concept)
Run: One cycle of prompt submission to AI response. Stored in the `runs` table. Referenced via `derived_from_run_id` etc.

Multi-Brand Architecture

Brand: The service delivery unit. Genre (BL / General Fiction / Fanworks) x Age Rating (All-ages / R18) = 6 brands. Settings managed via `brand_config` JSONB.; Tenant (DB table name)
Preamble: The `system_prompt` parameter built for LLM API calls. 3-layer structure: Layer 1 `system_base` (universal) + Layer 2 `brand_context` (genre-specific) + Layer 3 `content_guidelines` (age-rating-specific).
content_guidelines: Preamble Layer 3 key. Age-rating-specific expression guidelines (policies for sexual/violent content, etc.).; Content Policy (different scope)

Generation & Verification Pipelines

Interaction Mode: The PromptPane UI input mode. Two values: `chat` / `scene`. Former `rewrite` and `adaptation` modes have been moved to Artifacts View operations.
Deliberation Pipeline: FR-18. Runs 5 algorithms in series (peer_review / iterative_refinement / pruning / theme_metaphor / implicit_expression) to improve output quality. Default cost multiplier: 21x.; Deliberation Mode
Idea Palette: FR-17. A 9-stage pipeline (P0-P8) that generates creative ideas using 6 facets and a predicate system (30 common + 18 BL-specific predicates).; Idea Seed Generation (old name)
Consistency Verification (FR-08): Two-pass verification: Near-Range (sliding window over consecutive scenes) + Long-Range (MapReduce-style 6-category entity extraction). Detects contradictions in characters, equipment, locations, etc.; Integrity Check (old name)
Anchor Audit (FR-19): Focused verification around user-specified anchor scenes. Uses Prereq Checklists and Coverage Ledgers for intensive integrity checking.
Finding: An issue detected by Consistency Verification (FR-08). Classified by severity S1-S4. Scope: NEAR or LONG.; Issue (in FR-19 context)
Ono Engine: FR-24. A post-processor that expands `<ono>` tags into onomatopoeia text using tension-driven Markov chains. No LLM involved — pure application logic.; Onomatopoeia Engine
Adaptation: FR-07. Rewrites existing text based on setting information. 4 components: TriggerDetector -> ElementProtection -> AdaptationStrategy -> AdaptationProcess.

R18 Pipeline Components

ActionStep: The output of Phase 1 (structure estimation + dictionary search + random selection). A "translation layer" structure containing search keys for all 3 dictionaries.
ActionChain: ActionStep[] + global context (rating, register, scene_duration_hint). The action sequence for one scene.
HintPack: The output of Phase 3 (hint assembly). Integrates materials from 3 dictionaries per ActionStep. Injected into LLM prompts.; RefinedHintPack (after Phase 4)
Three-Axis Status Model: Soma (body) x Psyche (mind) x Meta (self-awareness) 3D state space. The "gap" between axes (body-leading / mind-leading / awareness-delay / full-axis divergence) is the design core that breaks stereotypes.
Anti-Blueprint Policy: Excludes 5 domains (pharmacology, trauma mechanics, physical restraint, chemical reactions, human physical limits) from verification checks. This is "exclusion of verification criteria" — not "injection of unrealistic modifiers".

Billing & Credits

Credit: 1cr = 1,000 output characters or 4,000 input characters. Plans: Lite (60/mo) / Standard (500/mo) / Pro (3,000/mo). 12-month expiry, FIFO consumption.; Cost (JPY amount)Budget (UI component)
Creative Lab: Conversational AI feature for drafting fragments and creative consultation. Free within daily limits, credit-based beyond. Formerly "Free Chat" (renamed via ADR-049).; Free Chat (old name)

LLM Architecture

LLMClient: Abstract LLM client replacing CohereClient (ADR-045). Determines model family (Claude / Qwen) based on `brand_config.content_policy.age_rating` and dispatches to the appropriate provider.; CohereClient (old name)
LLM Gateway: Abstraction layer between LLMClient and Provider Adapters. Responsibilities: provider selection, fallback control, test mode scenario provider substitution.
Model Family: All-ages = Claude (Haiku/Sonnet/Opus), R18 = Qwen 3.5 (9B/35B/122B 3-tier Elastic). Determined by `brand_config.content_policy.age_rating`.
Model Tier: Task-based model size selection: `light` (Creative Lab) / `standard` (scene generation) / `heavy` (deliberation mode). Selected by performance/cost criteria.
TTE (Text Transformation Engine): 77-process text estrangement transformation system. Each process is identified by an ore name (e.g., Quartz, Amethyst, Granite). 5 processing types: Type A-E.; Ore Process
UPP (Unified Progress Protocol): WebSocket-based protocol for displaying progress of long-running pipelines. Uses `ProgressEvent` types and renders via the `ProgressOverlay` component.