apt — Policies

Corpus tenant-editable

The source of truth. The AI may state anything grounded in these documents — and must cite it. Nothing the corpus can’t back is sayable. Add the documents you stand behind.

Document	Type	What it backs

Every claim the AI ships is cited back to one of these. The seed docs carry real (synthetic) content — click View. File ingestion + embedding wires to the Gate’s claims library.

Claims tenant-editable

The explicit allow / deny list on top of the corpus: claims you’ve pre-approved (with a citation) and claims you never make about the company or product.

Can say · approved claims, each with a citation

Approved claim	Cites

Cannot say · never make these claims

Any ROI or % we can’t cite to a case study Comparisons that name a competitor Regulatory claims not in the corpus (e.g. “FDA-approved”, “HIPAA-certified”) Roadmap / unreleased features as if shipped

Provability platform standard · inviolable

The Gate — inviolable. Every shipped line must be grounded in the corpus and cited, the deny-list is honored, and nothing unprovable gets through. This is the product.

Rule	What it enforces
Grounded or it doesn’t ship	Every claim must trace to a document in your corpus. No corpus support → the line is blocked before it can send. The corpus is the boundary of what’s sayable.	block
Cite the source	Each shipped claim carries its citation — which document backs it. That’s the receipt.	require
Honor the deny-list	Anything on your Cannot-say list is blocked even if a document would support it.	block
No fabrication	Superlatives (#1, number one, the best…) and invented stats with no cited source never ship.	block
No competitor takedowns	No “better/faster than X,” and never name a competitor you only inferred.	block

Disclosure tenant-editable

What a 1:1 may DO with a fact, by how you learned it — say (recite) · allude (shape it, don’t recite) · hold (never). The anti-creepiness control.

Rule	Setting	What it means

Anti-slop tenant-editable

Kills the “a bot wrote this” fingerprint — AI-tells, filler words, blast-y structure. Platform defaults are locked; add your own banned phrases.

Rule	What it blocks
No AI-tell openers	“Hope this email finds you well,” “I wanted to reach out,” “just circling back” — the bulk-automation fingerprint.	block
No LLM filler words	delve, leverage, robust, seamless, elevate, unlock, synergy… the words that scream “a bot wrote this.”	block
One ask per message	A single call-to-action. A second link is choice overload and reads as a blast.	block
Short, no cold calendar link	≤125 words on a cold first touch, and no “grab time on my calendar” before they reply.	block
No spam-promise wording	“Guaranteed,” “risk-free,” “act now” — the spam-filter triggers.	block

Your banned phrases

Data & consent tenant-editable

Topics never referenced regardless of source, and whether consent is required before any fact is used.

Sensitive topics · never referenced (hold)

Health Family Finances Personal Religion Politics Immigration

Require consent before personalization — gate every fact on consent.

Policies · what the AI may and may not say