Intelligence and review §1
Seven rungs, and a hard line across the middle of them
The blueprint calls this distinction “one of the best safeguards in the earlier BB2G concept” and says it “should remain in the schema, APIs, AI prompts and UI.” It is in the schema here, which is what makes it enforceable rather than decorative: every event carries its rung, and the ledger refuses to accept a machine-authored record that claims a rung it may not claim.
3 rungs a machine may assert · 4 that require a named,
attributed human. Read from LADDER in assets/lockchain.mjs at build time.
The rung most systems get wrong is ALLEGATION
A machine may record DATA. A machine may record an OBSERVATION, because a person or a sensor recorded seeing something. A machine may raise an ANOMALY, because a rule fired and the record names the rule rather than a cause.
A machine may not make an ALLEGATION. An allegation is an assertion by a person that something improper occurred, and it belongs to the person who made it. Software that generates its own allegations has invented an accuser who cannot be questioned, cannot be cross-examined and cannot be held responsible for being wrong. So ALLEGATION sits on the human side of the line even though it is below ANOMALY on the ladder — the ladder is ordered by evidentiary weight, not by who may write it.
EVIDENCE, FINDING and CONCLUSION are further up and equally closed to machine authorship. A finding is a reviewer’s reasoned determination, attributed and dated. A conclusion may carry legal weight. Neither is a thing a rule engine is entitled to produce.
The blueprint’s own examples
Three things the software is permitted to say
“Expected seal value differs from observed value. Review required.”
“Custody documentation contains a 43-minute undocumented interval. Review required.”
“Multiple sites reported the same printer-error code within 20 minutes. Correlation detected.”
Each one names a measured discrepancy and stops. A differing seal value. An undocumented interval of a stated length. The same error code at more than one site inside a stated window. None of them says why.
Five it must never say automatically
Fraud.
Criminal conspiracy.
Election tampering.
Intentional suppression.
Malfeasance.
The blueprint’s reason, in its own words: “Those are conclusions requiring evidence and, in some contexts, legal determinations.”
The distinction the schema keeps
A person can be agitated, make a threat or violate a rule without software attempting to diagnose why.
The ban is on machine authorship, not on the words
This is the part that is easy to get backwards, and getting it backwards produces a system nobody can use. A human reviewer may absolutely write fraud into a FINDING. A county attorney may write it into a legal conclusion. An investigator may allege it. Those are exactly the people the ladder reserves the upper rungs for.
What the platform refuses is a machine writing it on its own — a rule engine, a correlation job, a copilot, a template, an alert. The guard runs on machine-authored narrative text only. It never edits, softens or blocks a human’s words: a review whose language has been quietly adjusted by software is no longer that reviewer’s finding.
The guarded vocabulary
14 terms. If a machine-authored narrative contains one, the append is refused and the error names which word it refused, so the refusal can be read and argued with rather than merely obeyed.
fraud fraudulent criminal conspiracy conspiracy election tampering tampering intentional suppression suppression malfeasance rigged stolen election criminal illegally corrupt
A human-authored FINDING passes through untouched. The
flag on the event is machine_authored, and it is the flag — not the word — that decides
whether the guard runs.
What the refusal looks like
The ledger throws rather than silently rewriting, and the error is written to be read by an operator rather than a log parser:
// machine_authored: true, ladder_rung: 'FINDING'
LADDER_MACHINE_OVERREACH: a machine-authored record may not assert
FINDING — that rung requires a named human
// machine_authored: true, narrative contains a conclusion word
CONCLUSION_WORD_REFUSED: machine-authored narrative used "tampering"
— the system reports what a rule saw, never a cause or a culpabilityWhy this is a safeguard and not a gag. A discrepancy that is reported cleanly gets reviewed. A discrepancy that arrives pre-labelled as a crime gets argued about instead. The ladder does not make the system say less; it makes the system say the part it is actually entitled to say, and hand the rest to a person whose name goes on it.