Deliberation Pipeline

## Overview Raw LLM-generated prose tends toward flat, expository writing — telling rather than showing, relying on cliches, and lacking thematic depth. hakadoru.ai addresses this with a post-generation refinement system. **The Deliberation Pipeline is defined as a five-stage serial processing architecture where generated text undergoes peer review, iterative refinement, theme/metaphor enrichment, implicit expression conversion, and pruning — each performed by specialized LLM calls — to systematically elevate prose quality at approximately 21x the base generation cost.** This pipeline is optional and user-activated, designed for writers who prioritize prose quality and are willing to spend additional credits for measurably better output. ## Background The "generate and edit" workflow that most AI writing tools offer places the entire quality burden on the human author. The author generates text, reads it, identifies weaknesses, and manually prompts for revisions. This is time-consuming and depends on the author's ability to diagnose prose issues — a skill that varies widely. hakadoru.ai's Deliberation Pipeline automates the diagnostic and refinement cycle by decomposing prose quality improvement into five discrete, well-defined operations executed in sequence. Each operation has a specific mandate and cannot be skipped or reordered, ensuring consistent quality elevation. ## The Five Algorithms ### Stage 1: Peer Review The generated text is evaluated by multiple reviewer personas, each assessing different quality dimensions: narrative tension, character voice consistency, pacing, and sensory detail. Reviewers produce structured feedback (not prose rewrites) identifying specific weaknesses with line-level references. The feedback is formatted as a prioritized list of issues, each tagged with a severity level and a specific text span. This structured output enables Stage 2 to address issues systematically rather than attempting a holistic rewrite. ### Stage 2: Iterative Refinement Using the peer review feedback as input, the text is rewritten to address identified issues. This stage may execute multiple passes — typically 2-3 — until the reviewer feedback score stabilizes. Each pass focuses on the highest-priority issues from the remaining feedback. ### Stage 3: Theme and Metaphor Enrichment The refined text is analyzed for thematic opportunities. This stage identifies where abstract themes can be grounded in concrete imagery, where recurring motifs can be strengthened, and where metaphorical language can replace literal description. The enrichment respects the author's established tone and does not impose stylistic changes that conflict with the work's voice. ### Stage 4: Implicit Expression Conversion This stage mechanizes the "Show, Don't Tell" principle. It identifies passages where emotions, states, or character traits are stated explicitly ("She was angry") and converts them to implicit expression through action, dialogue, or sensory detail ("Her fingers whitened around the cup handle"). This is the stage that produces the most noticeable quality difference in typical LLM output. ### Stage 5: Pruning The final stage removes redundancy introduced by the enrichment and conversion stages. Prose tends to grow during refinement; pruning trims unnecessary adverbs, redundant descriptions, and over-explained subtext. The goal is net-neutral or reduced word count compared to the input, ensuring that quality gains do not come at the cost of bloat. Pruning operates with specific heuristics: sentences that restate what a previous action already implied are candidates for removal; adverb-heavy constructions are simplified; and passages where implicit expression in Stage 4 made an earlier explicit statement redundant have the explicit version removed. The pruner also checks for tonal consistency, ensuring that the cumulative edits from Stages 2-4 have not introduced jarring shifts in voice. ## Serial Execution Design The five stages execute in strict serial order. Each stage's output becomes the next stage's input. This is intentional: peer review must precede refinement (you cannot fix what you have not diagnosed), theme enrichment must follow basic quality fixes (enriching flawed prose wastes effort), implicit expression conversion must follow enrichment (to convert newly added explicit statements), and pruning must be last (to clean up all prior additions). Parallel execution was evaluated and rejected because intermediate stages produce text that is intentionally imperfect — enrichment may over-write, conversion may over-show — and subsequent stages compensate. Running stages in parallel would lose this corrective cascade. ## Cost Transparency The pipeline's approximately 21x cost multiplier relative to base generation is presented to the user before activation. The cost breakdown by stage is visible in the progress UI, and users can monitor credit consumption in real time. This transparency is essential: the pipeline is a premium feature, and users must make informed decisions about when its quality benefits justify the cost. The 21x multiplier breaks down approximately as follows: Stage 1 (peer review) accounts for roughly 4x due to multiple reviewer personas, Stage 2 (iterative refinement) accounts for 8-10x due to multi-pass rewriting, and Stages 3-5 each account for approximately 2-3x. The exact multiplier varies with text length and the number of refinement passes Stage 2 requires before convergence. ## Progress Visibility Each stage reports its progress through hakadoru.ai's Unified Progress Protocol (UPP). Authors see which stage is currently executing, the intermediate output at each stage boundary, and estimated time remaining. Intermediate outputs are particularly valuable — an author can observe how the text evolves through each transformation and develop an intuition for what the pipeline does, building trust in the process. ## Anti-Blueprint Policy For R18 brands, the Deliberation Pipeline integrates the Anti-Blueprint Policy, which ensures that generated intimate content avoids formulaic patterns. The peer review stage (Stage 1) includes an Anti-Blueprint reviewer persona that flags templatic structures, and the refinement stage (Stage 2) actively disrupts detected patterns while maintaining narrative coherence. The Anti-Blueprint Policy addresses a well-known weakness in LLM-generated intimate content: convergence toward a small set of narrative templates regardless of the characters, setting, or emotional context. By making template detection an explicit review criterion and pattern disruption an explicit refinement goal, the pipeline produces R18 content that is meaningfully varied across generations. ## When to Use the Pipeline The Deliberation Pipeline is not intended for every generation. It is most valuable for: - **Key scenes** — Climactic moments, emotional turning points, and scenes the author considers pivotal to the story - **Final drafts** — Scenes that have been structurally finalized and are ready for prose-level polish - **Quality benchmarking** — Running the pipeline on a sample scene to calibrate expectations for a new project For exploratory drafts, brainstorming sessions, or rapid iteration, standard single-pass generation is more cost-effective. The pipeline's value proposition is highest when applied to text that is structurally sound but needs prose-level elevation. ## Comparison with Other Approaches Most AI writing tools offer single-pass generation with optional manual re-prompting. Sudowrite's "Rewrite" feature performs single-pass revision without structured multi-stage refinement. ChatGPT-based workflows can approximate iterative refinement through conversation, but without systematic stage decomposition or specialized reviewer personas. hakadoru.ai's Deliberation Pipeline is distinguished by its formalized stage decomposition, its serial corrective cascade, and its transparent cost model. The pipeline treats prose quality as an engineering problem with defined inputs, operations, and measurable outputs — rather than as an emergent property of prompt crafting. The key architectural difference is that hakadoru.ai separates diagnosis (Stage 1) from treatment (Stages 2-5). Most tools conflate these steps, asking the LLM to simultaneously identify problems and fix them. Separating concerns allows each stage to specialize, producing more targeted improvements than a single-pass "make this better" prompt can achieve. ## Quality Metrics The pipeline's effectiveness is measurable through several proxy metrics: - **Explicit-to-implicit ratio** — The proportion of emotion/state descriptions that are shown through action versus stated directly, measured before and after pipeline execution - **Lexical diversity** — The type-token ratio of the output compared to the input, indicating whether the pipeline has enriched the vocabulary - **Redundancy score** — The semantic similarity between consecutive sentences, which should decrease after pruning - **Reviewer convergence** — The number of Stage 2 passes required before peer review feedback stabilizes, indicating the initial text's quality gap These metrics are tracked internally for pipeline performance monitoring but are not exposed to users, who evaluate quality through the more intuitive method of reading the before-and-after text. ## Conclusion The Deliberation Pipeline transforms LLM prose refinement from an ad-hoc conversational process into a structured, repeatable engineering pipeline. By decomposing quality improvement into five specialized stages — peer review, iterative refinement, theme enrichment, implicit expression, and pruning — the system produces consistently higher-quality output than single-pass generation. The 21x cost multiplier is substantial but transparent, enabling writers to make informed quality-versus-cost tradeoffs on a per-generation basis.