Research — AI Governance & Assurance
Requirements you can test.
Technical AI governance turns a requirement for an AI system into a control that can be tested, challenged, falsified and audited. AI assurance is the reproducible evidence that those controls actually hold under defined conditions.
The assurance cycle
From requirement to evidence.
INSPEKTORIUM does not ask whether a system is broadly “safe.” It investigates specific properties: a requirement becomes an invariant, the invariant becomes a control, and the control is only trusted once a test produces evidence under a defined fault.
01
Requirement
A property the system must hold.
02
Control
The mechanism that enforces it.
03
Fault
A deliberate way to break it.
04
Test
Run the fault against the control.
05
Evidence
What actually held, and where.
The full chain
Requirement → Technical invariant → Control → Failure condition → Test → Measurement → Pass / fail → Known limitation
Controls, concretely
Five requirements, five controls.
Each control is a specific, testable mechanism — not a policy statement.
Requirement
Each answer must trace to a retrievable source.
→ Control
Source binding
Claims are bound to source spans; unbound claims are rejected.
Requirement
Only authorized steps may run.
→ Control
Authorization gate
Every step passes an explicit authorization gate before it executes.
Requirement
An invalid state must not produce output.
→ Control
Fail-closed block
The workflow fails closed and blocks rather than emitting a best guess.
Requirement
Results must reflect the current inputs.
→ Control
Checkpoint rejection
Authenticated checkpoints reject stale or replayed results.
Requirement
Confidence must be well-formed to be usable.
→ Control
Calibration rejection
Malformed confidence is flagged as an invalid calibration state.
What AI assurance means here
It means
Collecting reproducible evidence about whether specified controls or properties of an AI workflow actually hold under defined conditions. Negative results and known failures are part of the research, not something to hide.
It does not mean
—
Certification
—
General AI safety
—
Regulatory approval
—
Guaranteed compliance
—
Guaranteed truth
—
Guaranteed robustness