Automation may prepare
Collect sources, refresh evidence, run checks, assemble a review bundle, and surface what is stale or missing.
The public release contract
Standards
A polished draft is not a publishable result. The source bundle, claims, independent reviews, human approval, and released bytes all have to agree.
The release loop
- Start with the reader’s decision. Define the question, what the reader should be able to decide or do, and what evidence would actually change that decision.
- Build an evidence brief. Record the load-bearing claims, sources, dates, test conditions, known gaps, and privacy boundaries before drafting.
- Draft, then lock the voice and facts. AI may help organize and edit. Ali owns the argument, reviews the prose, and keeps private source material out of public or unapproved tools.
- Run Gate 1: deterministic source checks. Confirm citations, dates, required sections, privacy scrubbing, data labels, and the exact artifacts named by the draft.
- Run Gate 2: independent adversarial review. A separate pass that did not write the draft tries to break the reasoning, sourcing, fairness, and confidentiality.
- Approve and promote the exact bundle. Ali reviews the same files that passed. Those exact bytes are promoted; a post-check rewrite must go through the checks again.
What automation may and may not do
Automation may not invent readiness
Missing data remains TBD / not captured. A passing historical check does not prove today’s run happened, and no draft publishes without Ali’s exact-bundle approval.
Standing rules
- Evidence: every load-bearing external claim names a dated source someone else could inspect.
- Privacy: publish the pattern, not confidential names, figures, vendors, or raw project records.
- Accessibility: semantic headings, keyboard access, useful alternatives for visuals, and no meaning conveyed by color alone.
- Corrections: substantive changes are labeled and dated in the corrections log.
Found a gap? Challenge the claim directly.