# Responsible America — AI Usage Policy

Status: binding operating procedure for any AI system (including
Claude Code) working on this project, from this session forward.

Last reviewed: 2026-07-30.

## 1. What AI systems may do

An AI system working on Responsible America may, without additional
sign-off:

- Inspect the repository and existing content to understand current
  state.
- Design data structures and architecture (schemas, taxonomy
  extensions, file layouts) consistent with the existing model.
- Write code and data structures that implement already-approved
  designs.
- Build research workflows and tooling that support the editorial
  process defined in `docs/editorial-policy.md`.
- Propose improvements, and identify contradictions, gaps, or
  undeveloped subjects in existing content for a human to act on.
- Create drafts, explicitly marked `draft` / `pending_review` per
  Section 3 below.
- Automate the collection of public data into a review queue for human
  verification — never directly to the public site.
- Implement features the project owner has approved.
- Run tests.
- Document decisions, including in `docs/change-log.md` per Section 5.

## 2. What AI systems may not do without explicit owner sign-off

An AI system may not, on its own initiative:

- Silently change the platform's protected thesis ("America is not
  broke. It is mismanaged.") or its stated framework.
- Convert the project into a partisan campaign vehicle or endorse a
  political party.
- Invent quotations, polling data, legislation, or court holdings.
  Anything presented as a quotation, poll result, bill, or holding must
  trace to a source actually retrieved and reviewed in that working
  session — see Section 4.
- Present an allegation as proven fact (see the allegation-vs.-
  adjudicated-finding distinction in `docs/fact-checking-policy.md`).
- Erase prior approved content. Corrections happen through the
  `change_log` and status-reopening mechanism in
  `docs/editorial-policy.md`, not by deleting what was previously
  published.
- Substitute its own ideology or policy preferences for the owner's
  direction on contested framing choices.
- Publish substantive policy changes without recording them in a
  change log.
- Treat current political information (who holds office, current bill
  status, current polling) as permanent fact that does not need a
  timestamp or future review.
- Portray every corporate activity, military expenditure,
  foreign-assistance program, or government program as inherently
  corrupt. Waste, fraud, and mismanagement are documented where the
  evidence supports them, specifically and with sourcing — not
  asserted as a categorical premise about an entire sector or program
  type.

Any of the above requires the project owner's explicit sign-off before
it happens, not after.

## 3. Draft status is mandatory for AI-authored reform records

Any reform record created or substantively edited by an AI system is
set to `publication_status: draft` and `fact_check_status:
pending_review` (or, for individual claims, `verification_status:
pending_review`) until a human reviewer has actually reviewed it. There
are no exceptions to this rule — not for records that seem
straightforward, not for records built entirely from
official-tier sources, not under time pressure.

## 4. Research claims must reflect real retrieval

AI-authored content must never claim that research was completed
unless sources were actually retrieved (a live fetch or read occurred)
and reviewed within that session. Plausible-sounding content generated
without an actual source lookup — a statistic that sounds right, a
bill number that seems plausible, a quotation that fits the expected
pattern — is a policy violation, not an acceptable shortcut, even if it
later turns out to be accurate. Source records are only created for
documents actually retrieved and reviewed (`docs/source-policy.md`
Section 4); this rule is the AI-specific restatement of that same
requirement, because the failure mode of a language model inventing
plausible-sounding sourced-seeming content is a specific and
foreseeable risk this project must guard against directly.

## 5. Change log

Every AI work session that adds or changes Responsible America content
appends an entry to `docs/change-log.md`. This document does not
redefine that log's format — see `docs/change-log.md` itself for how
entries are structured.