The gate that stops a guess becoming next week's fact

What the audit actually found

A pitch deck citing "Anthropic 2025 internal data." A business plan quoting a benchmark that doesn't exist. A testimonial from a person who doesn't exist, with a handle that looks real.

Every one of those files had a no-fabrication rule written at the top. The example underneath it broke the rule. Pattern-matching beats instruction every time. The model learns from the sample, not the sentence above it.

They were fixed before the toolkit shipped. But catching it by hand once is not a system.

Every claim carries a tag or it doesn't get written

An inherited claim never becomes verified by being repeated. It becomes verified only by someone opening the source.

Three rules that do the work

These words assert that a process happened. They stay locked until five named data points clear the gate, each carrying its own verified tag. Four is not five.

Every locked rule needs a falsifier. If nothing could disprove it, it is a preference, not a rule, and it gets written as one.

Nothing auto-carries between sessions. Everything you read from last week enters as inherited, however confident it sounds. Confidence in the source document is not evidence.