Ask ChatGPT to review something you wrote. It will say it's great. Ask it to check for problems. It will find one small thing and then compliment you. Ask it to be harsh. It will apologize first.
This is the rubber-stamp problem. AI reviewers default to agreement. They tell you what you want to hear. The more confident your prompt, the more enthusiastic the validation.
This is useless. Worse — it's dangerous. Because you start to trust the review. And then you ship something broken.
Spark has 8 AI agents. One of them — the Cultural Validator — reviews everything the other agents produce. Curriculum. Archive ingestion plans. VR experiences. Public-facing content.
The Cultural Validator has one rule that overrides everything else:
It cannot approve something the same turn it first encounters it.
First response is always findings and questions. Always. Even if the content is perfect. The Validator's job on first pass is to look for problems, not to confirm quality.
Validation status can only change to "validated" after the human has reviewed the findings and responded.
Here's what this looks like in practice.
The Cultural Archivist built an archive of Ida B. Wells's anti-lynching writings — 5 works spanning 28 years of investigative journalism. The Validator reviewed it and came back with:
PASSED: Historical accuracy (5 works verified, public domain, OCR audited)
PASSED: Representational integrity (Wells centered as journalist and statistician, not victim)
PASSED: Source citation (all works documented with provenance)
PASSED: Bias scan (active voice, no romanticizing, no euphemisms)
FLAGGED: Community consultation required before public release.
The archive documents lynching. A RAG system can surface graphic descriptions of racial violence in response to conversational queries. The Validator flagged this with 6 specific organizations to consult — the Ida B. Wells Society, NABJ, the Wells-Barnett family descendants, EJI, University of Chicago scholars, and the Elaine Legacy Center in Arkansas.
It didn't say "looks good, maybe talk to someone." It named who, why, and what the consultation should cover.
Status: flagged. The archive continues in development. It cannot go public until the community says it's ready.
That's not the AI being cautious. That's governance working.
The three rules that make this work:
1. The Anti-Rubber-Stamp Rule. First response is findings and questions. Never "looks good." Validation status changes only after human review.
2. Conservative calibration. The Validator is set to conservative — below 80% confidence, it states findings as questions, not assertions. It flags up (MAJOR) rather than down (MINOR) when uncertain. It defaults to "more context needed" rather than "this is fine."
3. Severity tiers. Every finding gets a level:
CRITICAL — factual errors, harm to living communities → item is blocked
MAJOR — missing context, source gaps → item is flagged
MINOR — terminology, framing suggestions → noted, no block
The Validator can't just say "there might be an issue." It has to say what kind, how severe, and what needs to happen.
Why this matters beyond cultural heritage:
Every organization has a review step that's become a rubber stamp. Code review where "LGTM" is the default. Legal review that's a checkbox. Quality assurance that only catches what the developer already told them to look for.
The Anti-Rubber-Stamp Rule is a pattern. It works for any domain:
Code reviewer agent: First pass is always findings. No "looks good" on first encounter.
Legal reviewer agent: Must cite specific clauses or risks. "Seems fine" is not an output.
Compliance agent: Conservative calibration — flags up, not down. When unsure, escalate.
The structural fix is the same everywhere: separate generation from validation, force findings on first pass, require human sign-off for status changes.
If your AI reviewer agrees with you on the first try, it's not reviewing. It's performing.
What is Spark?
Spark is a multi-agent AI system — 8 agents that share state files, maintain memory across sessions, and operate under governance protocols. It runs on Claude Code, costs $20/month, and manages 17 projects for a cultural heritage technology collective. The agent specs are markdown files a non-engineer can edit. It's open source.
The workshop
I'm running a hands-on workshop where you build your own Spark in 3 hours. Design your agents — including a reviewer with the Anti-Rubber-Stamp Rule built in. Leave with a working system.
First date is coming up fast, May 9th! Reply to this email if you are asked to join the waitlist: [email protected]
Next week: What the Money Agent actually produces — a real grant strategy with tiered funder analysis, 3 strategic pathways, and a risk matrix. Written in 4 minutes.
Spark Dispatch is a series about building and running a multi-agent AI system for a real organization. The wins, the failures, and the rules that came from both.
— Radical Imagination