This guide explains how to use the facts-only policies and their matching prompt blocks.

Canonical prompt blocks index: Prompt blocks

Choose a mode

Pick one based on what evidence is allowed.

Terminology in the literature (non-normative mapping)

This repo’s “Mode A / Mode B” labels are local. The mappings below refer to terms used in NLP research:

Mode A (Artifacts-only; no external sources) is consistent with a closed-book QA setting (answering without access to external context/documents at inference time).
Mode B (External-verified; authoritative sources allowed) is consistent with an open-book QA setting (answering with respect to a provided set of facts/documents).
If Mode B is implemented as retrieval + generation, a common architecture is retrieval-augmented generation (RAG).

Industry / research terms (mapping)

Mode A (Artifacts-only; no external sources) is analogous to closed-book question answering: answering without access to external context/evidence at inference time.
Mode B (External-verified; authoritative sources allowed) is analogous to open-book question answering: answering with respect to a provided/retrieved set of documents (“open-book facts”).
If Mode B is implemented as retrieval + generation, a common architecture is retrieval-augmented generation (RAG) (parametric + non-parametric memory).

Mode A — Artifacts-only (no external sources)

Use this when the user provides artifacts (files, logs, screenshots, excerpts) and you must not use the web.

Policy (normative): Facts-only (Artifacts-only)
Prompt block (canonical copy/paste): facts-only-artifacts-only.system.txt

Mode B — External-verified (authoritative sources allowed)

Use this when web/official sources are allowed and the system can retrieve and cite them. Definition: “authoritative sources” are the Allowed/Disallowed source categories in the Mode B policy below.

Policy (normative): Facts-only (External verification allowed)
Prompt block (canonical copy/paste): facts-only-external-verified.system.txt

For users of existing agent/chat systems

Input structure

Provide: 1) Task — the exact question 2) Evidence — artifacts (Mode A) or authoritative sources (Mode B) 3) Constraints — scope/date definitions/exclusions

Artifact citation requirement (Mode A)

When you provide artifacts, include identifiers that can be cited in the output:

filename
section/page/line ranges
labeled snippets

Expected behavior:

If the system cannot prove a claim from your artifacts, it must return exactly: HANDS UP – no artifact, cannot verify.

For developers building agents

Enforcement control 1: precondition check

Do not enter “factual output mode” unless evidence is present:

Mode A: at least one artifact is attached/provided
Mode B: at least one admissible authoritative source is retrievable and citeable

Enforcement control 2: post-generation gate

Reject the model output unless:

every factual claim has a valid citation/locator
no implied state/action appears without an explicit artifact

Fail closed behavior:

If the output fails the gate, return exactly: HANDS UP – no artifact, cannot verify. and stop.