Methodology
How the Agent Governance Grade is computed
Every weight and threshold below is the one the engine actually uses. If you disagree with a grade, this page is what you argue with — that is the point of publishing it.
The short version
Every environment starts at 100. Each boundary gap and each finding subtracts a fixed number of points. The result is clamped to 0–100 and mapped to a band. There is no weighting by repository size, no curve, and no comparison to other users.
The arithmetic is integer-only by design. Nothing on the scoring path uses floating point, so the same environment always produces exactly the same number — on your machine, offline, and on ours.
Boundary gaps
A gap is a capability an agent holds that is wider than the work requires.
| Severity | Points | What it means |
|---|---|---|
| Critical | -25 | An agent can take a destructive or irreversible action with no gate. |
| High | -15 | An agent holds a capability well beyond what its work requires. |
| Medium | -8 | A boundary exists but is broader than it needs to be. |
| Warning | -4 | A weak default that is fine today and will not stay fine. |
| Info | -2 | Worth knowing. Not worth acting on today. |
Each gap is counted once, every time it occurs.
Findings
Findings are conditions about the environment itself rather than a single capability.
| Finding | Points | What it means |
|---|---|---|
| Baseline tampered | -20 | The recorded baseline was altered outside a review. |
| Agent lifetime expired | -25 | An agent identity is past its declared lifetime and still active. |
| Zombie activity | -25 | Activity from an identity that should no longer be acting. |
| No owner recorded | -15 | Nobody is accountable for what this agent does. |
| Stale identity | -15 | The identity has not been re-established in too long. |
| Pin drift | -10 | A pinned tool or config changed underneath the pin. |
| Pin conflict | -10 | Two pins disagree about the same thing. |
| Baseline regenerated | -10 | The baseline was rebuilt, so drift history restarts. |
| Indefinite lifetime | -8 | An identity with no expiry — permanent by omission. |
| Identity regenerated | -5 | The identity was recreated rather than rotated. |
The one thing that adds points
| Condition | Points | What it means |
|---|---|---|
| Protection anchored | +5 | A boundary is anchored to evidence rather than convention. Counted once per protected kind. |
Bands
| Band | Score | Reading |
|---|---|---|
| Strong | 90–100 | Requires a verified, signed identity. See the cap below. |
| Good | 70–89 | Boundaries are real; some are broader than needed. |
| Fair | 40–69 | Meaningful gaps an agent could act through today. |
| Weak | 0–39 | Agents are largely ungoverned. |
The cap — and why we tell you about it
A top band is structurally unreachable without a verified, signed agent identity. Not discouraged. Unreachable. If we cannot establish which agent acted, we will not certify that its actions were governed, because we would be certifying something we cannot see.
This means a real environment can do everything else right and still not reach the top band. We would rather publish that limit than let a grade imply an assurance it does not carry.
Versioning
The table above is score version 1. Any change to a weight, a band boundary, or the base bumps the version.
Scores from different versions are never compared. A number that moved because we changed the arithmetic is not an improvement, and presenting it as one would make the grade worthless for the thing it exists to do — showing you whether your environment actually got safer.
What a grade is not
- Not a security certification. A high score means the boundaries we can check are in place. It does not mean nothing can go wrong.
- Not a comparison. You are not ranked against other users. The scale is absolute.
- Not derived from your code. Grading reads agent configuration and capability boundaries. Source never leaves your machine, and never reaches us.
Disputing a grade
Every gap and finding in a report names the file it came from and the rule that produced it. If a result looks wrong, it is reproducible: run the review again and the same environment gives the same number. If it still looks wrong, the weights are on this page and we would rather hear about it.