# Interaction — where the speakers stand on conversation

*02-08-2026 · the report behind the Interaction section of the [People page](index.html). Presenting (recorded narration) works today; this report is about TALKING WITH a speaker, and the knowledge behind her answers.*

---

**How a conversation works, in one line.** The visitor asks (typed or spoken), a hosted agent ([ElevenLabs](https://elevenlabs.io), our account) hears, finds the answer in a knowledge base we wrote, and speaks it back in the speaker's own cloned voice.

## What works today

- **Zalia (Comores).** The reference. Live conversation, text and voice, in French, on the [SARL conditions page](https://comores.eregistrations.org/conditions-immatriculation-sarl). Grounded on 5 citizen knowledge files plus a conduct charter; tested 9 for 9, including trick questions. Detail: [Zalia architecture](../../../2 - eR services/countries/Comores/knowledge/zalia-rag-architecture.md).
- **Grace (Lesotho).** Recorded narration only, on the [leaflets](https://businessregistrations.gov.ls). No conversation and no knowledge base — parked 19-07-2026 by Frank's verdict (the corpus is small; wake her when a public assistant is wanted). Detail: [lesotho-rag topic](../../../5 - Handovers/topics/lesotho-rag.md).
- **Valentina (AZFA).** The moving bust presents well at [azfa.eregistrations.dev/valentina-video](https://azfa.eregistrations.dev/valentina-video). Her conversation is fully plumbed — her own agent on our account, the fastest Spanish voice model, a pronunciation dictionary so "AZFA" is always said right, 13 knowledge files plus a charter — but the buttons were withdrawn 02-08-2026. No visitor can start a conversation yet. Detail: [ibero-freezones topic](../../../5 - Handovers/topics/ibero-freezones.md).

## What is missing for real conversation (Valentina)

1. **The entry buttons.** Put back the quiet text button on her greeting card and its bubble twin; both call the already-wired opener in [the page](../../../3 - Projects/ibero-freezones/valentina-video.html) (the code comment at line 2136 marks the spot).
2. **A full voice test on the live page.** A page opened from disk never captures the microphone; only the served page can prove voice works. The whole spoken round trip (ask, hear, answer) has never been run end to end.
3. **Interrupting her own answer.** Stopping the NARRATION now truly stops it (fixed 02-08-2026). Cutting her off mid-sentence during a LIVE answer, in voice mode, was never verified.
4. **The greeting in voice mode.** Her written greeting is long. Text mode hides the duplicate; in voice mode she may speak all of it before she listens. Length and first-answer delay need measuring.
5. **The face does not answer.** During a conversation the bust steps aside, idle (and disappears on phones). Her live answers are voice plus text bubbles — not the moving face. This is the real meaning of "talking, we don't know how to do it yet".
6. **Switching text to voice restarts the session.** The conversation history is lost when the visitor changes mode mid-dialogue.
7. **No context bridge.** The conversation does not know which section of the page the visitor was reading; the method's contextual hand-off is unused.
8. **Cost unmeasured, door open.** The agent is public with no domain lock, so anyone who finds its id can consume minutes. The price per conversation minute on the fast model was never measured (hosted minutes cost real money; the standing rule is to compare against a rented GPU).

## The knowledge behind the answers

**The idea, plainly.** Each speaker answers only from short fact files we wrote — one file per subject, every number traced to its source. Each file is labeled with WHO may see it (citizen, officer, internal). A small compiler sends only the citizen files to the agent, so internal notes can never leak: what is not in her index cannot be said. A conduct charter — what she may say, what she must never say — is compiled into her instructions. The whole method is the [voice-agent skill](../../../../.claude/skills/voice-agent/SKILL.md).

- **Zalia** has the complete build: labeled files, compiler, charter, and a guard that replays 12 witness questions (right prices, no leaks, refuses trick questions). But the guard's daily schedule is off — its scheduler is parked — so it only runs by hand. Still missing from her own plan: the server editor (edit a fact, press Publish, everything republishes) and the hub map.
- **Valentina** has 13 files plus a charter, built with the same method — but no guard, no witness-question set, and refresh is manual.
- **Grace** has no knowledge base; her raw material (the leaflets, legally verified against the Companies Act) is ready and waiting.

**Missing across the board:** a one-click refresh (change a fact → republish), a guard on a working schedule for every speaker, and witness-question coverage beyond Zalia.

## Improvement tracks (ranked)

1. **Prove Valentina's conversation.** Make talking with her actually work, then show it. Next task: on the live page, run the full voice conversation — microphone, answer, cut her off mid-sentence, greeting, delay — fix what fails, then restore the two buttons for Frank's verdict. Effort: about a day. Unlocks: the first speaker a visitor can talk WITH, and the demo for colleagues.
2. **Guard every speaker's answers.** A daily robot that re-asks the witness questions and flags any wrong or leaked answer. Next task: write Valentina's witness questions and trick probes, wire the existing guard to her, give it a schedule that runs. Effort: hours. Unlocks: answers that stay right as facts change — the condition for trusting any public speaker.
3. **One-click knowledge refresh.** A password-protected page where a colleague edits a fact and presses Publish. Next task: build the Zalia server editor (phase 2 of [her architecture](../../../2 - eR services/countries/Comores/knowledge/zalia-rag-architecture.md)); reuse it for every speaker. Effort: 1 to 2 days. Unlocks: non-developers keep the knowledge current.
4. **The face that answers.** Today the bust falls silent and steps aside while her voice answers. Next task: test real-time lip movement (our video factory's mouth engine on a rented GPU) against a live audio stream; keep the honest step-aside until it works. Effort: days, exploratory. Unlocks: the speaker herself visibly answering — the full promise.
5. **Wake Grace.** Parked by Frank's verdict; only when a public assistant is wanted. Next task then: run the voice-agent method on the leaflet corpus with the Grace voice. Effort: about a day. Unlocks: citizens on the Lesotho site ask instead of read.
6. **The sovereign bet (a labeled bet, not a task).** NVIDIA's [Nemotron VoiceChat](../../../../.claude/projects/-Users-unctad-Claude/memory/reference-nemotron-voicechat.md) early access: a self-hosted model that listens and speaks at the same time — you interrupt her as naturally as a person — answering in under a third of a second. Today: English only, no voice clone, so not for Valentina. Watch items: answer NVIDIA's contact asking for Spanish; later, try their Spanish-capable sibling pipeline on a rented GPU. Unlocks, if it lands: natural interruption a generation ahead, on our own machines, at centimes.