> ## Documentation Index
> Fetch the complete documentation index at: https://docs.thread.app/llms.txt
> Use this file to discover all available pages before exploring further.
# Compliance
> The always-on floor every guide runs on — consent, quiet hours, STOP, the gates every dial clears, no self-verification, escalation, and a complete log.
Every guide runs on an always-on floor. The floor is not a set of settings — it is enforced in the runtime itself, it cannot be turned off, and it holds no matter what a guide's instructions say. The guide page shows it as locked guardrail pills so everyone who edits a guide sees exactly what holds underneath it. Where you genuinely have a choice, it is named as one below.
The split is deliberate: the guide is trusted with judgment — what to say, when to nudge, when to check, when to bring a human in. **Code owns the physics** — consent, quiet hours, opt-outs, which number dials, and what closes.
## Consent before the first text
Enrolling a customer requires a consent attestation — one line covering every channel: *this customer agreed we may text and call this number about this outcome*. Who attested and when is recorded on the enrollment. There is no unconsented first touch. The gate runs before **every** send and **every** dial, not just the first, and refuses rather than assumes.
## Quiet hours, recipient-local
Overnight for the recipient means not sent. Quiet hours are 8pm–8am in the customer's local time:
* **Texts and emails are held**, and the guide's next wake moves into the recipient's open window — the message goes out at a decent hour, and the log shows the hold and the scheduled send.
* **Calls are refused, never deferred.** A queued dial that fires hours after the guide proposed it would ring out of context — so a call proposed in quiet hours simply does not happen, the refusal lands in the log, and the guide may propose again during open hours.
* **An unknown timezone blocks the send** rather than guessing. Under-sending is the safe failure.
## STOP ends it, instantly — everywhere
An opt-out reply (STOP, STOPALL, UNSUBSCRIBE, CANCEL, END, QUIT) is captured before the guide even wakes — the model never gets a vote:
* The suppression is **immediate and permanent**. It outlives the current outcome.
* It **spans every guide** in your organization — no other guide will contact that address either.
* It **spans channels**: a texted STOP also bars any proposed phone call to that number, and an email opt-out suppresses the address the same way.
## Every dial clears the same gates — whoever approved it
A guide's calls don't wait for a person by default. **Ask me before dialing**, on the guide's Phone calls channel, is off: the guide calls when its instructions call for it, the same way it already texts and emails on the same attestation. Turn it on and the guide pages you with who and why and dials nothing until you tap — on a surface showing the **full phone number**, which your tap echoes back so the server confirms a person saw the real destination. Those approvals are bound to the specific proposal and expire quickly; a stale approval dials nothing.
That preference is yours to set, and it decides only **who authorizes** a dial. The floor sits underneath it either way: every gate below re-runs against the customer's stored number in the moment before the phone rings, no matter who approved.
* **A STOP bars the dial**, across channels — a texted opt-out stops a call.
* **The consent attestation** must still cover this customer.
* **Quiet hours are refused, never queued**, and an unknown timezone blocks rather than guesses.
* **The number is never the guide's to choose.** It is resolved from the enrolled customer's stored phone at dial time; the guide has no field to put a number in, in either mode.
* **One approval, one dial** — and no guide calls the same customer again for several hours, so a call nobody answered is never retried.
So a guide asking for a call is not a guide getting one. A dial can be refused after the guide asks for it, and the refusal lands in the log with its reason. See [calls and approvals](/guides/voice-studio/calls-and-approvals).
## The guide never verifies its own work
Done closes only by citing a fresh read from a connected source, or by your explicit confirmation when no read tool is attached. Neither the guide's own word nor the customer's ever closes an outcome — the customer's "all set!" is a claim to verify, not evidence. This is checked in code at the moment of close, not left to the model's honesty. Full detail: [outcomes and verification](/guides/outcomes-and-verification).
## It pages your team when stuck
When a guide can't move forward — a question it can't answer, a claim it can't verify, a decision above its station — it pages the enrolling operator and waits. Escalating to a human is **never approval-gated**: the page is the point, so nothing stands between a stuck guide and your inbox. It pages once per stuck episode rather than retrying blindly, and the watch page grows one-tap actions to resolve it.
## Every step lands in the log
Every send, every hold, every refusal, every source read (including reads that found nothing), every escalation, every human tap, and the close itself is written to the per-customer log as it happens — with monospace receipts on every read. The [watch page](/guides/watch-page) renders that log in full, and the [five-stage funnel](/guides/outcomes-and-verification) is computed from it, never authored. If it isn't in the log, it didn't happen; if it is, you can show exactly when and why.
## AI disclosure on voice
Every published agent carries a disclosure rule its author cannot edit out: it is an automated assistant. Asked who they are speaking to or whether it's an AI, the agent says so plainly, in the call's own language. The rule leads the published prompt on every agent — a guide's instructions cannot override it, and no guardrail you write replaces it.
Whether that disclosure is also the **opening line** depends on the agent's welcome message: with no welcome message the agent opens with its pinned disclosure line; with a welcome message you wrote, that is what the caller hears first and the AI disclosure comes when asked.
## Recording
**Recording is a per-agent choice, and it is off unless someone turns it on.** Voice Studio carries one switch, **Recorded**, under the welcome message. Off, no audio of the call is stored on either leg — the written transcript is still kept, which is what the guide reads. On, the call's audio is stored and playable from the conversation sheet.
Turning it on is not just a storage decision, and the product treats it as one thing:
* **A recorded call says so in its first words.** If your welcome message already mentions recording — in your own words, however you phrase it — that stands, untouched. If it doesn't, publish adds the disclosure sentence to the opening line for you. There is no published, recording agent whose opening is silent about it.
* **The Studio offers the graceful version.** With **Recorded** on and a welcome that doesn't mention it, a dashed **Add "on a recorded line"** appears beside the switch and drops the phrase where your cursor is — so it weaves into your own sentence ("…calling on a recorded line to help get you scheduled — do you have a moment?") instead of being tacked on the end.
* **Asked outright, the agent answers honestly either way.** With recording on it confirms the call is recorded. With recording off it says the call is not being recorded and no audio is kept, and never claims otherwise.
* **A welcome can't claim what the switch doesn't do.** If **Recorded** is off but the opening line talks about recording, publish refuses and tells you to turn the switch on or take it out of the welcome. Nothing ships that tells a customer they're being recorded when they aren't.
Changing the switch reaches the two legs at different moments, which is worth knowing when you flip it: the call recording stops (or starts) on the very next call, while the voice vendor's own copy follows on the next **Publish**. Flip the switch and publish, and both agree.
Recording also has nothing to do with how long anything is kept. Conversations — transcripts and any audio alike — are retained for **30 days**.
The floor is a product guarantee, not legal advice. Messaging and calling regulations (such as TCPA in the US) vary by region and use case — confirm your consent language and calling practices with your own counsel.
## Related pages
Where the consent attestation is recorded.
How each channel behaves inside the floor.
The complete per-customer record.
Who approves a dial, and what code checks either way.