Fleet audit: agentic engineering discipline

Deterministic, scored reports on AI-agent development discipline across ryanportfolio's repos, most private. The tool and pipeline are public and deterministic; reports on private repos are reproducible by the owner from the pinned commit. Unflattering scores stay in.

What this score cannot see

E-1

Off-GitHub work is invisible

Work that leaves no GitHub artifact is invisible. Reviews run in a separate AI session, audits handed to another tool, and local test runs do not exist to this tool unless they were posted back to the pull request.

E-2

The solo blind spot

Solo accounts have a built-in blind spot: review credit requires a reviewer who is not the pull request's author. A solo owner who runs an independent review in a fresh session and posts the result from their own account gets no review-coverage credit for it. That can make a disciplined solo repo look weaker on review dimensions than it is.

E-3

Under-measurements, not corrections

These are under-measurements, not corrections. The tool states what it could not see instead of guessing.

D-1

Owner's declaration: reviews run off GitHubNot scored

Every repo on this board is solo. My reviews almost always run as handoff audits in separate AI sessions, which leave no GitHub artifact. The deterministic scorer cannot see those sessions and gives them no score credit, so the review dimensions here understate that practice. Where I have stated it for a repo, its report carries the statement as an attestation outside the scored evidence.

The skill behind that practice is public: /handoff-audit drafts a self-contained audit prompt for a zero-context session and charges it to falsify the work, not approve it.

Scoreboard

Repo	Score	Report
ryanportfolio/ryanportfolio	78 Strong	report →
ryanportfolio/Truenote	71.2 Developing	report →
ryanportfolio/Corewise.Academy	69 Developing	report →
ryanportfolio/AI-Firmware	68.6 Developing	report →
ryanportfolio/PixelSwarm	68.4 Developing	report →
ryanportfolio/Local-CPU-only-PTT	64.9 Developing	report →
ryanportfolio/range	56.6 Early	report →
ryanportfolio/Extract-Video-Wisdom	55.7 Early	report →
ryanportfolio/githelp	52.1 Early	report →

What this score means

S-1

Process discipline, not code quality

This score measures process discipline, not code quality. The tool never reads the code. It reads only GitHub metadata: commits, pull requests, reviews, check runs, and merge events. It scores only what is recorded on GitHub; discipline that leaves no artifact there earns nothing.

S-2

One question

It answers one question: when AI agents help write the code, what does this repo's history prove about the checks standing between a change and the main branch?

S-3

Scoring and grades

Each dimension is a 0 to 100 answer to one concrete question, or 'could not verify' when the evidence is missing. The overall score is the weighted average of the verified dimensions, and grades band it: 90+ Elite, 75+ Strong, 60+ Developing, 40+ Early, under 40 Ad-hoc.

How these scores are made

M-1

Built by the pipeline it documents

Every change to the audit tool itself flows: plan → agent build → independent fresh-context AI review → CI test+eval gate → owner-authorized merge. The repo's own PR history is the living demo: read it.