The Integrity Engine
We don't ask you to trust us. We show you the work. Every policy in this library passes through a four-stage verification pipeline before it reaches you — and every gate is open-source, auditable, and logged.
The Core Question: How Do We Trust It?
Three things make this library different from a random climate blog:
- The pipeline is public. Every rule the AI follows is written in
AGENTS.mdand enforced by open-source Python scripts anyone can read, run, or fork. - Humans are in the loop. AI agents flag problems; humans approve changes to the main branch. No AI can merge content autonomously.
- We use “Verification”, not “Truth”. Claims are either substantiated by cited evidence or flagged as unsubstantiated. There is no middle ground.
The Four Agents
Agent A — The Scientist
Role: Physical-reality check
Trigger: Any claim about emissions, health outcomes, or engineering specifications
Cross-references every factual claim against the cited evidence source. If the physics doesn't support the claim, the policy is flagged Unsubstantiated and cannot proceed until the citation is corrected or the claim is removed.
Agent B — The CFO
Role: Economic stress test
Trigger: Any mandate involving procurement timelines, construction, or cost estimates
Reviews supply-chain lead times and flags "Unfunded Mandates". Catches vague economic language — "strive to be cost-effective" → rejected; "shall not exceed $X per unit" → accepted. Timelines must reference a specific trigger event, not an absolute date.
Agent C — The Sleep Doctor
Role: Public-health impact analysis
Trigger: Any housing code, heat plan, or air-quality policy
Checks that every indoor-environment policy defines a Maximum Indoor Temperature for Nighttime Recovery. Policies that affect sleep and health without quantified thresholds are returned for revision — a number without a unit, or a unit without a source, fails this check.
Agent D — The Consistency Guardian
Role: Structure, citations, overlap, and enforcement language
Trigger: Every new or edited policy, plus a monthly full-library audit
Runs eight checks: frontmatter completeness, binding language, numeric threshold validity, overlap with existing policies, citation link health, geographic compatibility, readability, and an adversarial "red team" stress test. The script is open-source: scripts/consistency_guardian.py.
The Full Workflow
The diagram below shows the complete lifecycle of a policy — from first draft to published page.
_policies/
.bak file saved to archive/ before merge
Trust Mechanisms at Each Stage
Open-Source Scripts
Every automated check is implemented in public Python code in the scripts/ directory. You can read it, run it locally, and open an issue if you disagree with a rule.
Binding Language Enforcement
The Guardian script rejects vague verbs. "Encourage" → blocked. "Strive to" → blocked. Only "shall", "must", and "required to" pass. This is not optional — it is a hard filter in code.
Citation Link Auditing
Every URL in official_sources is checked for HTTP 200. Dead links are flagged immediately. When a link dies, the Wayback Machine archiver automatically proposes a replacement.
No Autonomous Publishing
AI agents produce reports; they do not merge PRs. A human maintainer must approve every change to the main branch. This is enforced by GitHub branch protection rules, not just policy.
Versioned Backups
Before any AI-assisted edit that changes >20% of a policy, a timestamped .bak file is saved to archive/. The commit message must record which agent performed the edit and why.
Public Audit Trail
Every change is a Git commit with a documented reason. The full history is publicly visible on GitHub. You can compare any version to any other version and see exactly what changed.
What the AI Does and Does Not Do
✅ AI does
- Reformat raw policy text into the standard structure
- Flag vague enforcement language for human correction
- Check whether citations are reachable and from authoritative domains
- Detect overlap with existing policies
- Run adversarial "red team" tests to find loopholes
- Generate a structured review report for human decision-making
🚫 AI does not
- Invent citations or fabricate data
- Publish or merge content autonomously
- Weaken enforcement language (doing so is a hard build failure)
- Remove official source citations
- Change numeric thresholds or safety factors without a flagged human review
- Access user data or send information to third-party services
Known Limitations
No automated system is perfect. Known failure modes include:
- A cited source may be misrepresented or summarized inaccurately — the agent checks the URL resolves, not the full content of every paper.
- Numeric thresholds may be transcribed incorrectly from tables or mixed-format documents.
- Legal terminology may not translate correctly across jurisdictions (Common Law vs. Civil Code).
- The adversarial stress test is only as strong as the patterns it has been trained to look for.
We treat every credible correction as a priority. Use the button below or open a GitHub Issue directly.
Run It Yourself
The full pipeline is reproducible locally:
# 1. Set up
python3 -m venv venv && source venv/bin/activate
pip install -r scripts/requirements.txt
# 2. Check all policies
python scripts/consistency_guardian.py --all
# 3. Check only policies you've changed
python scripts/consistency_guardian.py --changed
# 4. Auto-archive any dead links
python scripts/consistency_guardian.py --update-redirects --all
Source: scripts/consistency_guardian.py
Last verified: 2026-05-03 · Back to About →