Six agents.
One deterministic
pipeline.

"If the Research Agent detects a context gap, it should stop the pipeline and wait for Maya to fill it. A draft written with a known gap is worse than no draft at all."

Blocking pipeline. Research Agent asks questions. Maya must answer all before pipeline proceeds. Feels rigorous.

C-03 (no workflow disruption) is a binding constraint. A blocking clarification step transforms a 15-minute pipeline into a multi-session, asynchronous conversation that requires Maya to context-switch back. The value proposition collapses. Instead: gaps become [REQUIRES INPUT:] placeholders in the draft and HIGH-severity items in the Review UI. Maya resolves them in one focused review session — not scattered across multiple pipeline invocations. Non-blocking + explicit flagging is better UX than blocking + blocking.

Drafts produced with unresolved gaps contain placeholder text rather than real content. Maya must fill these in Review. This is the correct behaviour — a placeholder is honest about what it doesn't know. An inferred value is dishonest about what it does.

"LLM-reported confidence scores are notoriously unreliable. A model that's hallucinating will still report high confidence. This is false XAI."

Compute confidence externally from schema completeness (% fields populated), input length, gap count, and placeholder count. Fully deterministic. No LLM involvement.

External heuristic confidence scores are more reliable in a narrow sense but less informative in a useful sense. A document with all fields populated can still be wrong if the LLM made a plausible but incorrect inference. The LLM's self-reported confidence, combined with explicit uncertainty enumeration in the XAI card, gives Maya something actionable: not just "confidence = 0.82" but "I'm uncertain about these two specific things." The uncertainties[] field in the XAI card is where the real value is — confidence score is the headline, not the story. Maya reads both. We accept the theoretical unreliability of self-reported confidence because the uncertainties field provides the granular corrective.

Confidence scores may not be perfectly calibrated. This is disclosed in the Glossary. The Review UI labels them "agent-estimated confidence" to set expectations. The mandatory human gate means an incorrectly confident agent output is still reviewed before publishing.

"80 curated rules is a cherry-picked subset. A writer who passes the compliance check might still violate dozens of guide rules you didn't encode. You're creating false confidence in compliance."

Encode all 200+ verifiable rules from the full guide. More comprehensive, harder to criticise for omission.

The 80 rules were selected by impact, not convenience. They cover every rule that is (a) binary-checkable without deep semantic understanding, and (b) responsible for >90% of the style violations found in a sample audit of 50 real developer docs. The remaining guide content is either stylistic judgment (where a rule cannot be binary) or highly context-dependent (where an LLM check would produce too many false positives to be trusted). A compliance check that produces false positives destroys writer trust faster than one that has a defined, disclosed scope. The UI discloses the 80-rule scope explicitly — Maya knows she is getting the high-impact subset, not the full guide.

~20% of guide violations will not be caught by the automated check. The UI documents this. Maya is expected to exercise professional judgment on stylistic and context-dependent rules. The tool does not replace her editorial judgment — it automates the verifiable rules so she can focus on the non-automatable ones.

"Blocking export is paternalistic. Maya is a professional. She should be able to export with known violations if she chooses. A warning is enough."

Warning-only. Export enabled always. HIGH violations displayed prominently but not enforced. Maya decides.

P-02 (human gate is non-bypassable) is a binding principle, not a UX preference. The architecture's value proposition is that documents produced by the pipeline are compliant before they leave it. If export is available before HIGH items are actioned, the pipeline's compliance guarantee evaporates. A writer under sprint pressure will export and intend to fix it later — and then they won't. The gate is not paternalistic; it is the architectural integrity of the compliance pipeline. "Ignore with reason" is the escape hatch: Maya can dismiss a HIGH violation with a one-sentence reason, which is logged in the session record. That is not blocking; it is requiring a conscious decision rather than an accidental omission.

Maya must interact with every HIGH item before exporting. For a document with many HIGH violations, this extends the review time. This is the correct trade-off: review time is a feature, not a bug. The pipeline saves time on drafting; the writer invests that time in review. The net is still a significant improvement over the As-Is state.