CPLite

For publishers and author platforms

Authorial control for LLM book companions.

Stop your AI chat surface from reducing canonical texts to generic wellness slogans. CPLite's three-layer control stack holds companions to the author's voice, register, and argument — at 85 % mean conformity across a 9-book live collection, rising to ~95 % on matched author intent.

Request pilot access How it works →

The default behaviour of LLMs on canonical texts is drift.

Ask a stock GPT‑4o about Meditations and it replies in the register of a wellness coach. Ask about The Art of War and it reaches for MBA business-strategy vocabulary. Ask about Nathan the Wise and it reduces the play to an interfaith-harmony slogan that Lessing's text refuses. The model is not malfunctioning; it is returning its highest-training-mass register for each topic.

For a publisher, this is an editorial problem. Your readers are engaging with a voice that is not the author's, and not the book's. CPLite is the control stack that puts the author back in charge of that voice.

The three-layer control stack

Every companion response passes through three stacked layers of authorial control, independently verifiable and versioned per book.

  1. 1

    Hard rules

    Five imperative prohibitions, universal across every book, prevent drift into self‑help, wellness, motivational, or interfaith‑slogan register regardless of how the reader phrases their question.

  2. 2

    Per‑book governance

    The author's red‑line, tone, MAY / MUST‑NOT rules, and text‑grounded closing‑question directive (BGCQ) are compiled into the system prompt on every turn. Authors revise these through a structured governance interface, not free-form prompt editing.

  3. 3

    Author-approved exemplars

    Up to three Q→A demonstrations per book, authored by the rights holder, anchor the companion's register, length, and stance on register‑sensitive questions. In‑context learning pulls the model towards the exemplars at generation time.

Measured outcomes

~95 %

Mean authorial conformity on questions with a matched seeded exemplar.

~85 %

Mean conformity across a 9-book live collection, averaged over matched and novel probes.

~40 %

Baseline conformity from an unconfigured stock model on the same probe set. CPLite closes that gap.

Conformity scored against a fixed rubric: forbidden‑phrase suppression, chapter/scene citation rate, closing‑question text‑grounding. Full methodology in the accompanying hermeneutics paper.

Research foundation

CPLite implements the Canon Pack framework introduced in:

  • Vasse, R.B. (2026). The Canon Pack: A structured framework for authorial control over LLM‑based companion reading of canonical texts. Zenodo preprint. 10.5281/zenodo.19548105
  • Vasse, R.B. (2026). Governing Generative Interpretation: Authorial Control in LLM‑Based Reading of Canonical Texts. Submitted to LOGOS.

The platform behind CPLite — Living Literature — runs the reference implementation over a 9-book live subscription collection. living-literature.org

Pilot access

CPLite is currently in closed pilot with a small number of publisher partners. We are opening additional pilot slots for Q2 2026.

Pilot partnerships include onboarding for up to three backlist titles, author-governance training, and a measurement dashboard showing per-book conformity over the pilot period.

Email to request a pilot slot

API endpoint: https://cplite.living-literature.org/api/
Full technical documentation is shared under NDA after initial contact.