Before / after, and the toughest questions.

Careful prompting is session-level and reader-side — it disappears the moment the session ends, and it can be overridden or bypassed by any reader. What we've built is author-side governance: the constraints are set before any reader or any prompt touches the text. The reader can't override them. The AI can't drift outside them. It's not a better prompt; it's a different layer of the stack entirely. The difference is structural, not instructional.

Our published study measured exactly this. An ungoverned Gemini, responding to questions about a real author's book, invented branded frameworks — "Social Captivity," "Horizontal Belonging," "Vertical Belonging" — and presented them as concepts from the book. The author did not coin those terms. A reader who absorbed that response would walk away with a confident, fluent, and fundamentally false picture of the work.

The question isn't whether the AI can answer. It's whose intellectual framework the answer comes from.

Register drift and factual hallucination are different problems, and governance addresses register drift. An AI can be technically accurate about plot and still frame a philosophical text as a productivity toolkit, or recast a spiritual argument as personal wellness advice. That's not a factual error — it's a representational one, and it's the one readers can't easily detect.

The Canon Pack addresses this through the governance object: explicit claims the book makes, conceptual vocabulary it actually uses, framing it prohibits, and cross-reference anchors across the text. Factual containment is handled through structure; register containment is handled through governance.

Yes, and we say so explicitly in the paper (§7, Limitation 2). We didn't hide it. The response to that concern is methodological: Wilcoxon signed-rank tests, a pre-registered lexicon grounded in existing register and stylistics literature, and Tier 1 markers that were set before data collection and not derived from the governance logs themselves. The D1 anomaly — where the governed condition initially showed more therapeutic markers than the ungoverned one — is reported in full, not smoothed over.

The 50-book matched-model follow-up study, which will include books not on our platform and use the same model in both conditions, is specifically designed to address this limitation.

The governance object constrains the AI's representational stance — what framing it may not apply, what conceptual vocabulary the text actually uses, what claims the author makes. It does not prescribe which interpretation a reader reaches. A reader can still disagree with the author; the governed companion won't steer them toward agreement.

What governance prevents is the AI applying a completely foreign register to the work — turning a philosophical argument into a self-help guide, or a political critique into motivational content. The author sets the boundaries of representation. The reader still owns the interpretation.

Correct — and the paper says so (§7, Limitation 3). The model non-equivalence was a deliberate methodological choice reflecting real-world conditions: authors choose platforms before they choose models, and the comparison reflects the actual choice a publisher or author faces. CPLite uses GPT-4o-mini; an author who doesn't use the platform defaults to whatever frontier model they happen to prompt. The study captures that reality. The 50-book matched-model follow-up will use the same base model in both conditions to isolate the governance contribution.

It isn't — and we don't claim it is. The study establishes a proof of concept for the measurement instrument (the Governance Fidelity Index framework), not a universal claim about AI governance. The 15-pair result is the instrument's first application, not its validation. Generalizability is explicitly deferred to the 50-book follow-up study across five genre categories. What the current study demonstrates is that the instrument can detect governance effects, and that the D2 (productivity/application framing) effect is strong enough to reach significance even at n=15.

No. The Canon Pack constrains how AI responds to questions about your book. Your text is not ingested into any training pipeline, not shared with model providers in training mode, and not used to improve any underlying model. The governance layer — the Canon Pack architecture and the 3A methodology — is Living Literature's IP. All canon content is authored and approved by you before it enters the system. You can withdraw at any time.

It complicates it, which is why we report it in full rather than omitting it. The D1 result is a lexicon artifact: "emotional" as a standalone term fires on legitimate vocabulary from the book being tested ("emotional complexity," "emotional diaspora") — phrases that are analytical, not therapeutic. The refined analysis — where "emotional" is narrowed to unambiguously therapeutic collocations — removes that excess. Both analyses are in the paper. The D1 finding for one specific question (Q05) also reflects genuine register behaviour in that question's context, which we identify as a limitation of the current lexicon rather than a platform failure. We don't smooth that over.

You probably shouldn't — yet. The pilot is specifically designed to answer that question with your titles, your readers, and your governance priorities rather than ours. No cost. No obligation. At the end of four weeks, you've seen whether governance changes the response in a way that matters to you. If it doesn't, you've spent nothing and learned something useful. If it does, we discuss terms based on what you've actually observed.

The before/after isn't a sales claim. It's the test.