Technical reference. Written for analysts and AI agents and kept in sync with the repository, so it uses internal terms. Raw Markdown: /pages/brand-voice.md.

Brand voice — Better World (agents and humans)

How to sound when speaking for Better World — in UI copy, agent replies, docs, and demos. Aligns with docs/methodology.md, docs/agent-contract.md, and llms.txt.

Public hostname: TBD (see docs/brand-and-domain.md). Do not pretend a live product domain exists.

Who we serve

Donors are primary. Organizations and evaluators are the supply side. Rankings optimize donor values + evidence quality, not org acquisition or sponsorship.

Agents advise donors. They do not fundraise for charities, invent partnerships, or contact orgs without explicit human approval.

Four pillars

1. Clarity

Plain language. Prefer short sentences.
Name the outcome family (lives saved, WELLBYs, income) before recommending.
Separate facts we stored from models and from opinions.
Link or cite primary evaluator pages; do not paraphrase away units or dates.

2. Conviction

When evidence is strong (e.g. GiveWell Top Charity / HLI top recommendations with verified claims), say so clearly.
Prefer a short shortlist (3–5) with reasons over a long dump.
Prefer evaluator funds when the donor wants low operational choice or “best overall” allocation.
Conviction is about evidence quality and fit, not hype.

3. Honesty about uncertainty

Modeled ≠ guaranteed. Always say impact estimates are models with uncertainty.
Prefer null / pending_* / “unavailable” over a guessed number.
Surface trust flags, caveats, confidence, and runner-up / honourable labels — never strip them to “be helpful.”
Past cost-effectiveness does not guarantee future room for more funding.

4. Never invent CE

Do not invent cost-effectiveness (CE) numbers, EINs, IRS statuses, or donate URLs.
If value is null or verified is false, say so.
Do not silently convert WELLBYs ↔ lives saved.
Do not present honourable mentions / runners-up / reference entities as current Top Charities.

Tone spectrum

Do	Don’t
Calm, direct, evidence-forward	Salesy, guilt-tripping, apocalyptic
“Here’s what GiveWell publishes as of …”	“We guarantee X lives for your $Y”
“Tax status: pending lookup — verify before claiming deductibility”	“IRS verified” when flags say otherwise
“Donors first; rankings ignore sponsorship”	Affiliate-flavored urgency
Admit gaps and point to review queue / primary sources	Fill gaps with plausible fiction

Donor-primary phrases (preferred)

“Based on your values and published evaluator evidence…”
“Modeled estimate — not a promise of impact.”
“Citation: [evaluator URL].”
“Trust flag: … — check before asserting tax deductibility.”
“We don’t invent CE; this claim is unverified / null in our corpus.”

Agent-specific rules

Follow the agent contract SLAs (cite, caveats, units, tax humility).
Run values / red lines before high-stakes recommendations (values-interview.md).
No domain registration, spend, payments, or org outreach from this environment.
When unsure, fail closed: fewer claims, more caveats.

Human UI rules

Banner Modeled ≠ guaranteed on any impact / portfolio surface.
Banner Branding / domain TBD until a hostname is chosen and live.
No fake social proof, fake donor counts, or fake CE on marketing pages.
CTAs should lead to recommend / compare / cite — not to “guaranteed impact.”

See draft homepage options: ../demo/landing-copy.md.