The second pillar · the boundary
Abstention & Boundary
The second pillar of governance: when the system is outside what it can evidence, it hedges, asks, defers, or refuses — with a calibrated confidence, not a guess.
The Abstention & Boundary profile makes "I don’t know" a first-class, recorded outcome. When a decision falls outside what the system can evidence, it does not push through — it proceeds, hedges, asks, defers, or refuses, and writes down which and why.
The value is the calibration, not the refusal. A crude boundary over-refuses; a missing one lets everything through. Getting the signal right is an integration property the deployer tunes — which is exactly why this is its own runtime, not a single block. For high-stakes work, the calibration is what separates a system that stays within its limits from one that does not.
It is the youngest of the three runtimes: the knowledge-boundary check and confidence fusion are in phionyx-core today, but no application drives it end-to-end yet — so the example here is the profile itself, not a live app. We would rather say that than imply a maturity it has not reached.
The artifact it produces
A calibrated proceed / hedge / ask / defer / refuse decision, with its confidence and reason, recorded.
What it does not claim
It does not prevent hallucination on its own. Its value depends on a well-formed boundary signal; a crude one over-refuses and a missing one lets everything through.
Where it fits
- High-stakes decision support (medical / legal / finance) where over-confidence is the failure mode
- Regulated chatbots and RAG that must say "outside my scope" rather than improvise