Why is the AI so confident when it's wrong?
The Confidence Contract is the explicit commitment that every Aegis recommendation carries a defensible confidence score, a source trail, and a refusal pathway when the evidence is insufficient. The operator-facing version of Truth Architecture.
The unsigned contract behind most AI output.
Open any consumer LLM and ask it for a financial forecast, a competitor analysis, or a hiring recommendation. The answer comes back articulate, structured, and confident. There's no score on it. There's no 'here's how sure I am.' The user reads it the same whether the model is 95% confident or 35% confident.
This is the implicit contract operators are signing without reading: each recommendation looks equally certain, so the operator has to do the calibration work themselves: usually under-skilled and time-pressured.
The Confidence Contract is Aegis Boardroom's explicit replacement. Aegis recommendations are designed to carry a documented confidence category, the sources that drove it, and a stated refusal pathway when the evidence does not support a recommendation.
What the score means.
I Know: high confidence, advisor-validated.
Multiple corroborating primary sources. Recent. Domain-specific. The recommendation is defensible to a board with the source trail attached. Most strategic-level Aegis recommendations target this state.
I Think: moderate confidence, recommendation only.
Single primary source, or multiple secondary sources. Directionally correct but should be validated before high-stakes execution. The output flags exactly which validation step is needed.
I'm Inferring: low confidence, observation only.
Inferred from indirect signals, model knowledge, or single secondary sources. Offered for ideation, not execution. The output reads 'consider this hypothesis' instead of 'do this.'
I Don't Know: insufficient evidence, inquiry mode.
The system declines to produce a recommendation. The output describes the gap, names what would be needed to close it, and identifies who would be the right source. The most valuable output in regulated, high-stakes, or novel situations.
Frequently asked questions.
Who scores the recommendation: the model or a human?
Both, by design. The Confidence Contract is built into how Aegis configures the model: the model surfaces its evidence and uncertainty as part of producing the answer, and a human advisor reviews high-stakes outputs before the recommendation lands. The advisor doesn't generate the state from scratch; they confirm or override it.
Can I see an example score?
Yes. Sample reports include confidence states against each major finding, and the AI Readiness Assessment itself returns a state on your readiness category.
What if I disagree with the state?
The override is documented. Operators can override a recommendation, and the override is recorded with the operator's reasoning alongside the engagement record. Six months later, both signals are available when calibrating future work.
How is this different from Truth Architecture?
Truth Architecture is the underlying technical framework. The Confidence Contract is the explicit operator-facing commitment. The contract is what you read. The architecture is how it gets honored.