AI Integrity Policy

Version 1.0.0 · Effective 2026-08-12 · Applies to all research, published writeups, and site content

1. What this policy is

Optimized Workflow is an AI-systems consultancy, and our published research — Published research, benchmarks, and recommendations — is produced with AI assistance by design. This policy states the standards that assistance must meet. It is modeled on the integrity policies of institutions that publish AI-assisted research at scale: Brookings (research independence and integrity), Harvard Business Review (editorial standards), the ACL Rolling Review (AI-assistant disclosure), and AP (newsroom standards for AI-assisted content).

2. Human accountability

• Every piece is published under a named human's responsibility. AI systems draft, analyze, and summarize; humans decide what is true, what gets published, and what gets corrected.
• Humans are accountable for accuracy, integrity, sourcing, and originality — no "the AI said so" defense. Editorial decisions are made by humans, not by models.
• AI output must be validated by a human expert before inclusion in anything we publish. Validation means checking claims against primary sources and, where a number is published, re-deriving or re-running it.

3. How AI is used in our pipeline

Our research-to-publication pipeline has ten stages. AI assists in these:

• Gather — AI agents collect and snapshot primary sources (tool-call budgets, source-quality rules, primary-over-secondary).
• Outline & draft — AI proposes structure and drafts sections from the gathered sources.
• Verify — a dedicated citation/fact-check stage places and verifies sources; no claim passes without a source.
• Edit — AI and humans edit for tone, neutrality, and clarity.

These stages are always human:

• Scope & plan approval — a human approves the research question before the heavy run.
• The Publish Gate — a 15-item checklist (claims match scope, limitations stated, methodology re-runnable, numbers explained, sources licensed, AI use disclosed) that must pass before anything ships. "No" is allowed if justified — the gate enforces disclosure, not compliance.
• Review — technical review (claims and numbers) and editorial review (style and clarity) by named reviewers.
• Publish & corrections — humans hit publish and own the corrections log.

4. Disclosure standards

• Every published piece carries a disclosure block stating that AI assistance was used, what it did, and that a human verified the content — the model-card metadata principle applied to reports.
• Per-report AI disclosure (ACL ARR standard): each published report states that AI was used in research, coding, and writing, and records the work log on request (method, dates, model versions).
• AP floor: AI-generated output is reviewed and edited by humans before publication, and AI use is disclosed when it plays a material role.
• No undisclosed AI-generated images. Any synthetic image or asset is labeled.

5. Independence

• No paid endorsements. Published research is self-funded ($0 in every report to date) and measures our own workloads. We do not accept payment for coverage, and no vendor reviews or approves a report before publication.
• Relationships are disclosed, not hidden. Each report states whether we build or operate the systems measured and whether we hold positions in or have paid relationships with any company mentioned — including whether we may seek compensation from them in future.
• Track record is public. We publish how past recommendations performed. A recommendation history we cannot show is a recommendation we should not make.
• No ghosting. If we stop covering a research line, we publish a final report explaining why, comparable in scope to the work that came before.

6. Data & training

• We do not train models on client submissions or on our published content without consent. Client data used in a report is described (what it is, how it was collected, how it was labeled), never dumped.
• Report data is FAIR where possible: findable, accessible, interoperable, reusable. Datasets and gates are available on request with stated terms and licenses; we prefer open artifacts hosted on persistent platforms with stable identifiers.
• Every number has a source. If we cannot source a number, we do not publish it.

7. Corrections

• Corrections are logged, not buried. Post-publication changes appear as inline "Edit <date>: …" notes inside the document, and the report version bumps.
• Version history is preserved. Reports are versioned and dated; superseded versions remain identifiable. This policy itself is versioned below.
• We correct promptly and credit the finder. A found error is a gift; we record who found it when they want that.

8. Policy version history

9. Contact

Questions, corrections, or requests about this policy: (603) 748-4982 or via the workflow review form.