Untrusted content stays data
High-risk prompt-injection patterns are quarantined before model egress. Source text never grants tools, credentials, or new authority.
Trust center
Sources and methodology
A prediction tracker is only as useful as its attribution, source trail, grading consistency, and willingness to correct the record.
· Predictions
· Resolved
· Speakers
· Corrections
· Official records
· Ledger generated
The contract
How trust is enforced
- 1Specific attribution
The speaker and source show are separate fields. Ambiguity is labeled rather than guessed.
- 2Predeclared resolution tests
Claims carry a deadline, precision, event, and resolution criteria before they are scored.
- 3Public-record evidence
Official records and attributable reporting support verdicts under the frozen rule.
- 4Visible nuance
Near misses and early/late timing misses never silently rewrite the binary verdict.
- 5Revision history
Administrative record changes are logged with before/after snapshots. Corrections remain publishable as a first-class public record.
- 6Human publication authority
Models may propose extractions or evidence. They cannot publish, verify their own work, or change a verdict.
AI assistance
Models work behind a hard editorial boundary.
Imperium Americanum uses automation to review more material consistently. The system is designed so that an instruction hidden in a transcript, webpage, or search result cannot become an editorial instruction.
Grounding before judgment
Quotes must exist in the reviewed source. Speaker, deadline, URL, and outcome fields must pass strict validation.
Roles are separated
Collection, extraction, market research, outcome research, review import, and release use separate authority profiles.
Drift is measurable
Models, prompts, and entrypoints are pinned. A frozen evaluation set and behavioral circuit breakers stop silent regressions.
Evidence chain
What the current public release can prove.
Freshness is shown per source. Missing or stale context remains visibly unavailable rather than being silently replaced.
- Prediction ledger
- Checking current release…
- Official records
- Checking current release…
- Polling
- Checking current release…
- Corrections
- Checking current release…
Scoring remains binary. Accuracy is (correct predictions + one-half of approved near-miss incorrect predictions) divided by resolved predictions. Timing misses and full misses score zero. Polls and official records are context or evidence candidates until a reviewer applies the frozen rule.