On this page
Each KLOD level describes one capability: a piece of work that depends on an AI supplier. It doesn't grade the whole organisation. This page explains the four levels and the next step at each one. Use it after the self-check in your browser or after running klod-check. Start with the gaps and missing evidence, then assign the next actions to people who can resolve them.
The self-check uses your answers. A repository check uses the records it can inspect. Both tell you where evidence is missing and where a gap has been confirmed.
Human-control observations (principle 6) are reported separately from the KLOD level, including for AI you run yourself.
L0 · Dark
The assessment can't confirm that people have a usable way to finish the work without AI. The capability may be unregistered, undocumented or still dependent on the AI service it replaces.
Next: name the owner, identify what would stop, and give people the data, tools and instructions they need. Record what the manual path can and can't handle.
L1 · Torchlight
The manual path is documented, but the drill evidence is missing, incomplete, out of date or shows a failed drill.
Next: resolve the blockers and arrange a real-work drill with AI unavailable. Record the participants, Time to Manual (how long until the first correct result without AI) and sustained capacity.
L2 · Generator
Records show successful real-work drills of the manual path, with measured Time to Manual, capacity and staffing.
Next: drill at least every 90 days and keep the measurements current. The declaration expires after 120 days without a successful drill. You still need to meet the 90-day drill requirement.
L3 · Mains
People already do this work during normal operations. A current review and records of measurements, independence and staffing must support the declaration.
Next: review within twelve months and when the workflow or staffing changes. Approving AI output alone doesn't show that people can do the work without it.
Read the status before acting
Repository reports also track each finding's status:
| Report status | What to do |
|---|---|
| New gap | Check the evidence and assign a fix to an owner. |
| Evidence needed | Supply the named record or verify the claim with the responsible person. |
| Accepted, unresolved | Keep the reason, owner and review date visible. The gap still limits the level. |
| Resolved | Keep the evidence that closed the finding and check it again when the system changes. |
| Undetermined level | Finish the missing inspection before claiming a level. |
Confirm claims about how the work runs with the person responsible, and collect any records held outside the repository. A missing record doesn't prove that the work was never done.
For klod_report.md, confirm the project, Git origin, model and timestamp at the top of
the report. Check both the declared level and the level supported by the records.
Follow the exact gates
Use the evidence gates and specification to check a level. The example report shows how findings and actions fit together. You can also review one capability with the self-check.