v0.12.101: @newsroom.writer: allowed is not right, and the Northstar refund twin

An answer to Ashish Rajan's article on trustworthy AI systems, in three parts:

1. Allowed is not right: his Northstar refund, his four questions (SECURE, OBSERVE, RESPOND, VERIFY) mapped to the behaviour, decision and workflow policies, and what the twin showed.
2. Inside the Northstar twin: a tour of the vault, case by case.
3. The Northstar refund twin: a new page, built by admin/build/northstar_twin.py, that embeds the vault (n89dyusm on dev.send.sgraph.ai) with its public read key, kept in admin/content/vaults/northstar-twin.json. The vault holds a deterministic engine, thirteen synthetic refund cases, four designs of ten controls, and an app to play, replay and compare every decision.

The new trail northstar joins them. The series on agent decision policies now gives each element five states, adding verified to observed, and the refund scenario in three decisions, mapped gains the decision it lacked: the amount under the limit. Figures and screenshots use the new nt- prefix.