The problem: fluency is not the bar
A model will write any number of plausible episodes. The question is whether an episode uses only the characters and facts valid at this point in this work, whether it follows the structure of the reference bestsellers, and whether its length and style are within bounds. This engine does not ask the model to judge that.
Distillation packets: extracting structure from the database
Distillation here means selecting and structuring database material, not compressing model weights. It has four layers. D1 flow: the distribution of same-stage reference episodes (reward, growth, cost, relationships, new characters, hooks, endings) and author chapter lengths. D2 episode: Q&A-based answers plus the database beat plan (order, weight, length). D3 scene: per-beat reference scene blueprints (roles, movement, state before and after, expression blocks). D4 sentence: reference text per blueprint block. The engine also decides which author's style applies — a recorded failure, where rotating sentence material from different works per beat produced an 'eight-author average' style, is the basis of that decision.
Database gates: no material, no model call
Four gates went in on September 30, 2026. Without material, answer 'no material' and the range to extend instead of calling the model. Fill the evidence column only with database-checked items. Let bestseller-trend data decide the skeleton of a proposal. Let the database decide beats and reference scenes before composition. The context snapshot selects only introduced characters, published facts and active rules, never treating later material as current knowledge.
Checks and verdict: nothing below the bar is published
A composition passes structured-output checks, a length range, a style-binding ratio and detection of repeated narrative endings. Attempts are bounded and each attempt's problems and usage are recorded. On October 13, 2026 episode 19 failed its first composition for three consecutive identical endings; episode 39 on October 26 failed for the same reason; a new work's second episode on November 2 was rejected at 4,439 characters (4,800–5,800 required) and a style binding of 70 of 102 paragraphs (80% required). Review-guided patching runs on a separate path, keeps the original failure and passes the same checks again.
Model independence
Generation runs through an injected external LLM call. BLUEURBAN's engine owns material selection, context and structure, and result checks; the packet and check layers stay unchanged when the base model changes. The technology described here is the control layer around generation, not the base model.
The service opened on this engine: Taste Edition
The repository began on September 28, 2026 and the first unattended run on the server was October 11. Episode 18 was published at 6,505 characters on October 13 (second composition passed, five close-reading checks); episodes 23, 24 and 25 all passed on the first composition on October 14; on October 19, the third day of one-a-day, episode 32 was published at 7,372 characters; November 1 marked the fifth day of a two-a-day trial. Meanwhile the service gained its portal and member access (10-03), a payment-review deployment (10-07), operations statistics (10-11), six MCP read tools (10-23) and pre-launch fixes (11-02). One serial reached episode 49 on November 3 and one original reached episode 33 on October 29. 1,138 commits in forty days.