Grounded Gate Protocol · v4.5 · M365 Copilot

Nothing passesunverified.

An anti-hallucination framework that tags every AI claim, runs reputation risk checks, and keeps you in control before anything reaches a client.

Section 01

AI sounds authoritative. It isn't always right.

Large language models produce fluent, confident text. They also fabricate statistics, misattribute quotes, and present assumptions as facts — with exactly the same tone as verified information. In professional settings, that is a liability.

Client email

Wrong delivery date cited as confirmed

Executive briefing

Market figure from an outdated blog post

Regulatory submission

Invented citation in a compliance section

Section 02

GGP makes uncertainty visible before it ships.

Grounded Gate Protocol is a set of instructions that forces Copilot to be explicit about what it knows, what it is assuming, and what it cannot find. Every response comes back as a marked draft — not a finished document. You decide what passes.

Verified.

Every CONFIRMED claim cites source, tier, location, and date.

Labelled.

Inferences and gaps are flagged inline — never buried.

Decided by you.

Clean output is only produced after your approval.

Section 03

Three phases. Every time.

🧯

Phase 01

Verify

Tags every factual claim before it reaches you.

—🟢 CONFIRMED — verified with Tier 1–3 source + full citation
—🔴 GAP — missing information declared openly, never filled
—🟡 INFERENCE — assumption made explicit for your decision
—⚪ UNVERIFIED — could not verify or source quality too low
—⚠️ RISK — reputational, legal, or credibility concern flagged

🛡️

Phase 02

Protect

Runs an 8-dimension risk check on every draft.

—Misinterpretation — could this be weaponised out of context?
—Credibility — are all claims sourced and defensible?
—Legal — any liability exposure?
—Reputation — does this align with your brand values?
—Data Accuracy — verified statistics with methodology?
—Hostile Reader — any exploitable weaknesses?
—Screenshot test — comfortable if shared without context?
—CEO check — would leadership approve without question?

✅

Phase 03

Deliver

Marked draft first. Clean output only after you decide.

—Every flagged item shows a unique code (I1, G1, R1…)
—Inline decision options with strategic context for each
—Decision summary table — reply with your choices
—Full reference audit with tier, URL, date for every source
—Clean output produced only after your confirmation

Section 04

Every claim gets a flag.

Instead of a polished output that hides uncertainty, GGP embeds visual markers inline. You see what is confirmed, what is assumed, and what is missing — right in the text.

🟢

CONFIRMED

Fact verified with Tier 1–3 source. Full citation included.

🔴

GAP

Critical information is missing. Declared openly — never hidden.

🟡

INFERENCE

An assumption the AI is making. You decide to accept or reject it.

⚪

UNVERIFIED

Could not verify, or source quality is too low.

⚠️

RISK

Reputational, legal, or credibility concern. Review before sending.

Section 05

Not all citations are equal.

GGP classifies every source by quality tier. In Copilot, your SharePoint and OneDrive documents count as Tier 1 when officially approved. A blog post cited as fact is flagged differently from a government filing.

Tier	Trust	Sources
T1 Gold — Authoritative	High	SharePoint/OneDrive docs, SEC filings, government data, peer-reviewed journals, audited financials
T2 Solid — Reputable	Good	FT, WSJ, Reuters, BBC, Gartner, McKinsey, Forrester, official industry reports
T3 General — Cross-reference	Moderate	General news, expert blogs, company marketing, Wikipedia (context only)
T4 Caution — Never CONFIRMED	Low	Forums, social media, anonymous blogs, commercial bias, outdated >2yr

Section 06

See it in action.

This is what a GGP-processed email looks like before you send it. Every non-confirmed item shows decision options inline — you choose, then GGP delivers the clean version.

GGP Marked Draftemail · Fast Mode

Subject: Project update — delivery and next steps

Hi [Client Name],

The data migration phase completed on schedule 🟢 CONFIRMED and all validation checks passed with no critical failures.

Performance testing shows a 34% reduction in response time 🟡 INFERENCE: I1 compared to the baseline.

I1 — Derived from internal test results not validated against production load. If production differs, this claim may undermine confidence at go-live.

a) Acceptb) Rewrite as GAPc) Provide sourced) Remove

🟢 1🟡 1GGP 7/8 · awaiting I1 decision

Ready?

Set up GGP for Copilot
in under 10 minutes.

Copy three fields into Copilot Agent Builder. Upload four knowledge files. Publish.