Every check in the mechanical layer, with its pillar, their point ceiling, and the effort it takes to close them. Published and versioned, because a score you can't audit is an opinion with decimals.
Documentation
25· 4 checks
Agent instructions
14· 5 checks
Executability
26· 7 checks
History
20· 5 checks
Structure
15· 5 checks
26 checks · Rubric version 1.16
The 26 checks, grouped by what it costs to close them. Almost half the score sits in things you write once: an instructions file, a README, an .env.example.
These are the same checks and the same 100 points as the tables below, grouped by the Effort column they already carry. There is no new data here: there is another way to read the same data.
Write the map once, not in every prompt. Why it matters →
| Check | Points | Effort |
|---|---|---|
Documentation coverage1.1-doc-coverage | 15 | medium |
Documentation freshness1.2-doc-freshness | 5 | low |
Root README1.3-readme | 3 | low |
Navigable documentation1.4-doc-structure | 2 | low |
Leave the instructions where the agent looks. Why it matters →
| Check | Points | Effort |
|---|---|---|
Agent instructions2.1-agent-instructions | 4 | low |
Adapter consistency2.2-adapter-consistency | 3 | low |
Live documentation links2.3-doc-links | 3 | low |
Always-on context2.7-always-on-context | 2 | medium |
How to verify, told to the agent2.8-verify-loop | 2 | low |
Let the agent run and verify on its own. Why it matters →
| Check | Points | Effort |
|---|---|---|
Test volume2.4-tests | 3 | high |
Continuous integration2.5-ci | 4 | medium |
Test breadth2.6-test-breadth | 4 | high |
Reproducible install5.1-reproducible-install | 4 | low |
Discoverable commands5.2-discoverable-commands | 4 | low |
Declared environment variables5.3-env-example | 4 | low |
Strict verifier5.4-strictness | 3 | medium |
Look after the files that always break. Why it matters →
| Check | Points | Effort |
|---|---|---|
Risk hotspots3.1-hotspots | 7 | high |
Concentrated knowledge3.2-knowledge-concentration | 5 | medium |
Abandoned areas3.3-zombie-areas | 3 | medium |
Commit traceability3.4-commit-traceability | 2 | low |
Documentation follows the work3.5-doc-drift | 3 | low |
Make it fit in a context window. Why it matters →
| Check | Points | Effort |
|---|---|---|
Giant files4.1-giant-files | 5 | high |
Committed dependencies4.2-vendored | 3 | low |
Tree navigability4.3-navigability | 4 | medium |
Libraries with the same role6.1-duplicate-role | 1 | medium |
Duplicate files6.2-duplicate-files | 2 | low |
Because an instrument you can't inspect isn't an instrument, it's a verdict. If you're going to let a number tell you your repository is half-ready, you have the right to see where every point comes from — and to argue with it. The rubric changes, which is why it carries a version.
A failed check returns its playbook — what to do and the criterion by which it closes — for your agent to run.
These checks measure presence. Whether what's there tells the truth is judged by the deep layer, with its own rubric.
Deep layer — judges substance