Research · White Paper
Verified Human Understanding as Cognitive Infrastructure
A shared accountability primitive for AI-assisted work and learning.
Tor Kringeland
Head of Research
Overview
AI made producing an artifact cheap, but understanding it stayed expensive — and a merged pull request or a submitted assignment no longer guarantees that the accountable person could explain the work. This paper defines the category Ninchi builds in: verified human understanding, an auditable record that a named person could explain the AI-assisted work they were accountable for, at the point of decision.
It walks through the verified-understanding event and the challenge loop, why the rubric-bounded grader is designed to be auditable rather than trusted by fiat, the transparent difficulty-weighted Ninchi Score in production today, and the proposed uncertainty-aware trust model and validation program that the next generation of the system is being built and tested against.
Shipped today
Generated, artifact-specific, timed challenges over GitHub, GitLab, and Bitbucket Cloud; hidden rubrics; threshold-based pass/fail scoring; the difficulty-weighted Ninchi Score; policy modes and opt-in enforcement; an attributable scored-event record per event; and org analytics over those records.
In development
Tamper-evident hash-chained records; confidence-based routing; human review, disputes, and overrides; a calibration program against human adjudication; and org-level concentration and bus-factor modeling.
Note on status: the paper separates what is implemented today from what is in development. Claims of calibration and prediction are stated as validation targets, not results.