The method

Sourced, verified, and signed by a person.

A method you can't inspect isn't a trust signal — so here's ours, in full. Every claim we publish was checked against a primary source by a person. We use AI to move faster on the legwork; the judgment stays human.

The pipeline

Nothing reaches a public page unread.

Monitor→Draft→Red-team→Verify — human→Publish

The first four steps lean on tools and move fast. The fifth is human, and it’s the one that matters: no source is published unread, and nothing auto-publishes — not a record row, not a claim built into a scenario, not an explainer. The judgment is a person’s.

Where the tools help

Four narrow jobs — none of them “decide what’s true.”

The model is a research assistant with a fast pen and no authority. Each job is bounded, and each output is a draft a human still has to clear.

Monitor

AI scans open sources — government documents, regulator dockets, reputable reporting — for new incidents and developments worth a closer look.

Draft

AI produces a first-pass summary of a candidate row or brief section, in plain language, with its claimed sources attached.

Red-team

A second prompt argues against the draft — flagging weak claims, missing context, and anything that reads stronger than the source supports.

Standardize

AI helps normalize records into the same dated, sourced, sector-tagged shape, so the dataset stays consistent and machine-citable.

Five standing commitments

The gate, in writing.

These are the rules the method runs on. They’re public so you can hold us to them — and so you can tell when we’ve broken one.

A published method

This page shows the pipeline, the gates, the data sources, and every human verification step. The method is the trust signal, so it's public.

A red-team fact-check harness

Before anything publishes, a second prompt argues against the claim; a human reviewer must resolve each objection on the record.

Citation-chasing as a hard gate

Every cited source is opened and read by a human. No model-supplied URL is ever published unverified — that's the gate nothing crosses.

Null results published

Quiet weeks are published as quiet. Wrong rows are retracted in the open, with a dated entry in the public corrections log.

A standing limitations section

Every product carries an honest note on what the method can and can't catch — the known failure modes, named below.

Mapped to a federal standard

We use NIST’s vocabulary on purpose.

The method maps to the NIST AI Risk Management Framework Generative AI Profile (NIST AI 600-1, July 2024), so the risks we guard against have names a reviewer already knows.

Confabulation

A model stating something false with confidence. Held by the human verification gate and the red-team harness.

Information integrity

Keeping the record accurate, dated, and traceable to a primary source — the whole point of the open tracker.

Human-AI configuration

Designing the human as the decision-maker, not the rubber stamp — adversarial review, never automation-by-default.

Limitations, named

What this method is built against.

These aren’t hypothetical. Every figure here was checked against its primary source before we published it — including correcting the numbers our own working drafts first carried. That’s the gate, applied to ourselves.

01Hallucination

Language models predict plausible text, not verified truth — so they still invent facts and sources, and some residual rate always remains. Each generation hallucinates less, but “less” isn’t “none,” and a neutral record can’t run on “usually right.” The human gate is built to hold whatever the current rate is — it doesn’t assume the model is bad, only that it isn’t perfect.

Early measurements (Stanford RegLab, 2024) put AI legal-research tools at 17–33%; the design assumption is that the failure shrinks but never fully disappears.

02Fabricated citations, in the wild

Court filings caught citing AI-fabricated cases had passed 1,400 worldwide as of June 2026, and the count keeps climbing. The failure is already real, not theoretical.

Damien Charlotin, “AI Hallucination Cases” database (updated continuously)

03Misgrounding

The most insidious error isn't a made-up source — it's citing a real source that doesn't actually say what the model claims. Only a human opening the source catches it.

Why citation-chasing is a hard gate, not a spot-check

04Automation bias

The documented human tendency to over-trust automated output. It's why the verification gate has to be adversarial review by design — a reviewer looking for the error, not a rubber stamp.

NIST AI 600-1 names this as a human-AI-configuration risk

When we get something wrong, we correct it in the open — in the record itself, rows are fixed in place, re-dated, or pulled, each against a named source.

Kick the tires

Found a claim that doesn’t check out?

That’s the whole point of publishing the method. Send us the source — we’ll review it, and if we’re wrong, the correction lands in the open with a date on it.

Email the lab →See the verified record

A line we keep in writing

Diplo Space, Inc. is a neutral, public-interest research and education lab. It uses open-source and public-domain data to translate space-governance regimes into navigable scenarios and exercises for non-specialist officials. It takes no policy positions, accepts no foreign-government funding, and is not affiliated with the U.S. Department of State, the Department of War, or any government agency.