INSPEKTORIUM
Technical AI Governance & AI Assurance Research
INSPEKTORIUM investigates how requirements for AI systems can be translated into technical controls that can be tested, challenged and audited. Scientific Claim Forensics is its flagship application.
The idea in one picture
Twenty artifacts pass a check. Switch between baseline processing and the full control layer to see what each one lets through.
Internal experiment, not independent validation. The counts are the recorded pilot result; which individual artifacts the baseline caught is illustrative.
The research question
How does a requirement for an AI system become something you can test?
INSPEKTORIUM does not claim that an AI system is generally safe. It examines specific properties of a workflow under defined conditions, and reports where the controls fail.
How the cycle works →
01
Requirement
02
Technical invariant
03
Control
04
Failure condition
05
Test
06
Measurement
07
Pass or fail
08
Known limitation
What exists today
Every capability carries its status.
Evidence-span validation
Answers are checked against the exact source spans they cite.
Implemented + tested
Artifact-integrity fault containment
Corrupted artifacts blocked in a controlled internal experiment.
Internal experiment
Semantic inference checking
Does the evidence warrant the inference, not just the citation?
Passed internally — not yet in public repo · Phase 8
General AI safety
Specific properties under defined conditions only.
Not claimed
The full status page →
Flagship application — Scientific Claim Forensics
Does the evidence actually warrant the claim?
A source can be genuine, correctly quoted, and still not support the conclusion drawn from it. Pick a leap and see which rung the evidence reaches, and which one the claim stands on.
“A blood marker fell after eight weeks of use.”
“Improves your health outcomes.”
The source supports the lowest rung. The claim sits on the top one. The missing step is: the marker predicts outcomes for patients, and nothing in this source shows it. A citation to this source does not close the gap.
Schematic examples. The wording is invented to show the shape of each inference gap; it is not drawn from a case file.
Internal experiment — not independent validation
What the first pilot showed, including what it did not.
24 / 24
No accuracy gain
All three conditions answered every question correctly. A ceiling effect: the pilot cannot show a benefit either way.
4 → 12
Of 12 corrupted artifacts blocked
Baseline blocked 4, the full control layer blocked all 12, and all 6 clean artifacts were accepted.
2 / 2
Wrong but valid outputs still passed
The controls check structure, not meaning. Semantic correctness is not shown, and general AI safety is not claimed.
Run the experiment →
Roadmap
A finite path to v1.0.
INSPEKTORIUM v1 is deliberately bounded. Everything past the final validation gate is a separate program — not an open-ended promise.
Now
Implemented + tested
Current foundation
Evidence binding, provenance and deterministic fingerprints.
Phase 8
Passed internally — not yet in public repo
Deep Forensics
Semantic and inferential integrity: does the evidence warrant the inference?
Phase 9
In progress
Disagreement & review
Independent second opinion and structured disagreement between reviewers.
Phase 10
Roadmap
Reporting & operational finish
A canonical audit trail and release-integrity workflow.
Gate
Planned
Final v1 validation gate
One consolidated validation before release.
v1.0
Release
The v1.0 release
The bounded release. Nothing follows it inside v1.
There is no Phase 11. Anything beyond v1.0 is v1.x, v2, or a separate research program.
Code & reproducibility
Two halves of the same work.
On GitHub
Code
Method
Tests
Reproducibility — the public v0.1 kernel
On this site
The research program
Explanation & Q&A
Experiments & evidence
Status & roadmap
github.com/LAKITALKS/Inspektorium
Not all internal research code is public yet. The public repository is the narrow v0.1 evidence-binding reference kernel — the starting point of the research, not the whole of it.
CONTACT
Research enquiries and methodological exchange.
For research enquiries, methodological exchange and institutional correspondence:
Please make contact before sending confidential or sensitive material.