Reasoning layer for edge agents — when should PicoClaw escalate vs act?

[bjornshomelab]

The Problem

PicoClaw agents on edge hardware face a question that
no current layer answers:

"Should I act on this — or stop and ask?"

Without a confidence gate, an agent on $10 hardware will:

• Act on uncertain inputs (false confidence)
• Generate plausible answers it treats as facts
• Reinforce errors in subsequent reasoning

This matters more on edge devices than anywhere else:
no fallback, no human in the loop by default.

A Minimal Solution

The core of a reasoning layer is just math:

conf := sigmoid(E_mean - D - 0.8*U_A - 0.6*M)

switch {
case conf >= 0.75: return PROCEED
case conf >= 0.50: return CAUTION
default:           return ESCALATE
}
Where:

E_mean = mean evidence strength across hypotheses
D = divergence between competing answers
U_A = assumption uncertainty
M = meta-uncertainty
Zero external dependencies. ~50 lines of Go.
Fits the 10MB constraint easily.

Reference Implementation
I've built this in Python as CognOS — a reasoning layer
for agentic AI that's been empirically tested across
frontier models (GPT-4o, Claude, Gemini, Mistral):

https://github.com/Applied-Ai-Philosophy/cognos

The research paper: https://doi.org/10.5281/zenodo.18731535

Proposal
A cognos-go package — or a built-in confidence gate
in PicoClaw — that lets edge agents know when to act,
when to ask, and when to escalate to a larger model.

Happy to contribute a Go port of the core confidence
engine if there's interest.

Is this something PicoClaw needs?