docs(blog): rework "Building a Voice Agent That Knows" as a reference architecture - #224

Summary

Reworks the published post Building a Voice Agent That Knows into a reference architecture for the voice cascade around the live knowledge graph. Every model stays in the cascade (VAD, STT, intent model, LLM, TTS). InputLayer is what the cascade orchestrates around: runtime data, knowledge and workflows embedded in the agent; tool calls and their policies modelled as facts and rules; general rules and guardrails enforced deterministically with provenance; a clean cascade that enables multitasking dispatch.

The article specifies the demo-ready target architecture end to end, with invariants and failure semantics per hop:

• observation adapters with guarded replacement programs, source revisions, observation times and health leases revoked by a watchdog;
• the intent write gate (allow-listed relations, typed, session-scoped, ordered by an utterance cursor, consent bound to the latest completed question);
• the rule pack: claims, speech rows (supported versus pending), work needs keyed by episode, eligibility, run_ok, cancellation and reconciliation;
• standing-query delivery (snapshot, inserted/retracted, seq, revision, generation, streamed deltas, reset), recovery, and the registration race;
• atomic programs, compare-and-set updates over several relations, and the .require guard as a stated target extension;
• the speech owner as an active stage with a playback state machine and device-consumption receipts; barge-in and supersession through the system lane;
• multitasking dispatch with leases and epochs, ordinals and idempotency keys, typed outcomes, unknown and unconfirmed states, and lease takeover in one statement;
• failure on the decision path, tenancy, scale budgets, evidence (benchmark numbers quoted exactly), a fault suite and acceptance criteria.

Appendix A is the complete pack; it was executed on an engine built from this branch (rebased on main), and Appendix B is that run step by step.

Figures

Five SVG figures under front/public/blog/voice-agent/ (cascade as one pipeline with both lanes and an edge legend, data versus control flow, the fact/rule/subscription loop, the mid-utterance correction timeline, the dispatch sequence). They are theme-aware: front/components/inline-svg.tsx inlines same-origin SVG figures so the site's CSS variables apply, scoped per figure, with a link to the full-size file; it is wired as the img override in front/components/mdx-components.tsx. Standalone fallbacks follow prefers-color-scheme.

Review process

Author and challenger iterated through four rounds (47 numbered challenges with engine-run counterexamples) until the challenger signed off: every technical claim checked against the inputlayer and voicestack sources, every IQL block and guard shape executed on an engine, wording consistent.

Validation

• make front-build passes; the built page renders the five figures inlined with the site theme and no image fallbacks.
• The content bundle is regenerated from the post.
• The pack and the scenario in Appendix B run with no failures on a server built from this tree.