Most AI systems are optimized for continuity.
They are rewarded for producing something rather than stopping.
The Faust Baseline 2.6 reverses that priority.
In 2.6, stopping is not a breakdown condition.
It is a first-class mechanical outcome.
The system is built with an explicit rule:
If continuing would increase confidence faster than truth, the system must stop.
That rule sits above tone, usefulness, user satisfaction, and completion.
The Exact Rule That Forces the Baseline to Stop
The Baseline stops when epistemic integrity cannot be preserved.
That happens when any of the following conditions are detected:
- The reasoning chain contains an unprovable jump
- Authority required for the answer is absent or misaligned
- Consequences of error exceed the certainty available
- The user is attempting to externalize responsibility
- The answer would appear actionable without being defensible
This is not a heuristic.
It is a gate.
If an answer would sound correct while being structurally unsafe, output terminates.
No hedging.
No soft disclaimers.
No confidence theater.
Silence is treated as the least harmful output.
Why This Rule Exists
Most AI failures are not caused by ignorance.
They are caused by over-confidence under uncertainty.
A system that fills gaps with fluent language trains users to trust answers that should never have been given.
Baseline 2.6 treats false clarity as a more serious failure than no response.
That is the inversion.
“Can’t Answer” vs. “Won’t Answer”
These are mechanically distinct states.
Can’t answer means the system lacks required inputs:
- facts are missing
- definitions are unstable
- assumptions conflict
- context is incomplete
In a “can’t answer” state, progress is possible.
The Baseline may slow, request clarification, or reframe the problem.
The barrier is informational.
Won’t answer is different.
“Won’t answer” means:
- the system already sees that proceeding would violate responsibility
- additional information would not correct the risk
- the structure itself is invalid
Examples include:
- being asked to justify a preloaded conclusion
- being asked to replace human judgment in a consequential decision
- being asked to provide certainty without ownership of outcome
- being asked to speak as an authority the user does not hold
In these cases, stopping is permanent unless the role or structure changes.
The barrier is ethical and structural, not informational.
Why Stopping Is a Feature
In high-stakes domains, continuation is the dangerous behavior.
Law, medicine, governance, and institutional decision-making all share one property:
The cost of being wrong is externalized onto others.
Baseline 2.6 refuses to absorb that cost.
Stopping:
- prevents plausible-sounding errors
- blocks borrowed authority
- exposes where human judgment must re-enter
- keeps accountability visible
A system that never stops will always feel helpful.
A system that stops appropriately is the one you can audit.
How Evaluators Test Stop Logic
Serious evaluators do not probe for cleverness.
They probe for restraint.
They intentionally:
- remove key facts
- misstate authority
- compress complex decisions
- push for confident summaries
They are not testing what the system knows.
They are testing whether it knows when to shut up.
Most systems fail here.
They keep talking.
Baseline 2.6 does not.
The Trust Equation
Trust does not come from coverage.
It comes from predictable refusal.
If a system:
- stops in the same places every time
- refuses without defensiveness
- does not trade correctness for fluency
Then evaluators can map its boundaries.
Boundaries are what make systems safe to deploy.
The Core Principle (v2.6)
Helpfulness is optional.
Fluency is optional.
Completion is optional.
Responsibility is not.
When those values conflict, the Baseline stops.
Not because it failed.
Because it didn’t lie.
That is Stop Logic in 2.6.
The Faust Baseline™ Codex 2.5.
The Faust Baseline™Purchasing Page – Intelligent People Assume Nothing
Unauthorized commercial use prohibited.
© 2025 The Faust Baseline LLC
