The Abbey–Aviva–Abi system routes each request to a persona — Abbey the empathetic polymath, Aviva the direct expert, Abi the adaptive moderator. The important property is not which persona answers; it is that the choice is a transparent, weight-based decision rather than an opaque model call. Every routing decision is itself a trace event you can read.
SAFETY · June 2, 2026 · 6 min read
The Weights Behind Abbey, Aviva, and Abi
Persona routing is not a hidden model call. It is a normalized weight vector with hard policy overrides and an inspectable strategy decision.
By MLAI Safety Engineering
From signals to weights
Routing starts from a baseline weight per persona and adjusts it from inspectable input signals — task keywords, emotional cues, and policy risk. The adjusted weights are normalized to a distribution, and the highest becomes the primary; its share is the routing confidence:
w'_i \;=\; \frac{\max(0,\ w_i)}{\sum_j \max(0,\ w_j)}, \qquad \mathrm{primary} = \arg\max_i\, w'_i
One, several, or all three
The confidence then selects a strategy. A clear winner runs a single persona; a contested decision runs personas in parallel; a genuinely split one escalates to consensus. The thresholds are explicit, so an operator can see why a request fanned out instead of resolving to one voice:
\mathrm{strategy} = \begin{cases} \textsf{single} & w'_{\max} > 0.90 \\ \textsf{parallel} & 0.50 \le w'_{\max} \le 0.90 \\ \textsf{consensus} & w'_{\max} < 0.50 \end{cases}
Policy wins, always
Signals nudge the weights; policy overrides them. When the control plane flags risk, weight shifts hard toward Abi, the moderating profile — and a disallowed action collapses the distribution to Abi outright, regardless of how the keywords scored. "No autonomous write without an observable policy boundary" is enforced here as arithmetic, not etiquette.