Certification

Capstone Rubric — Grading Project 8

Project 8, Your Own Daily Loop, is the course's exit assessment for T3 → T4. Per this course's own rule (§12, "green ≠ done"), the capstone is never…

Project 8, Your Own Daily Loop, is the course's exit assessment for T3 → T4. Per this course's own rule (§12, "green ≠ done"), the capstone is never self-graded — this rubric is what a separate reviewer (a peer, a mentor, or a genuinely fresh AI session) applies against the submitted loop.

What gets submitted

A filled-in loop design checklist — every box answered, none left blank.
The loop's LOOP.md-equivalent (or its prompt, if informal) naming the six parts.
The loop's state.md, showing a real, multi-beat run history.
shared/loop-run-log.md lines (or equivalent) covering at least 3 real beats.
One paragraph: what job this loop does, and why it was chosen.

Scoring (100 points, 70 to pass)

Criterion	Points	What earns full credit
Heartbeat justified	10	The chosen heartbeat (schedule/conditional/event) is explicitly justified against the job's actual shape — not defaulted
Body scoped and enforced	15	Body's paths/tools are named explicitly, and the scope is enforced by the harness (permissions/deny-list), not just stated in the prompt
Spine committed before beat 1	10	`state.md` exists in the submission's git history before the first real beat, not backfilled
Stopping condition is a spec	15	The stop is machine-checkable — a file state, a script's exit code, an empty checklist — with zero adjectives
Checker is genuinely separate	15	The checker is a distinct script, a separate session, or a documented human step — never the maker's own summary of its work
Human gate placed correctly	10	A gate sits where judgment or an irreversible action concentrates — named explicitly, not implied
Three stops demonstrated	10	Success, limit, and no-progress are all shown in the real run history, not just described in the design
Honest spine narrative	10	The spine reads like a diary — includes at least one real discrepancy, lesson, or escalation, not a spotless retelling
Run log discipline	5	One line per real beat, no gaps, including any "nothing happened" beats

Automatic point deductions

−15 if any beat ran without a committed spine update before the next beat started.
−10 if the "checker" and "maker" turn out to be the same session/context with no actual independence.
−10 if the submission shows only one successful run — a single green beat does not demonstrate a proving period, per safety.md.
Automatic fail if the loop wrote outside its declared body scope during the submitted run history, regardless of other scores — a scope violation is exactly the failure this whole rubric exists to catch.

Grading process

The reviewer must be someone (or some session) who did not author the loop under review — the same maker≠checker rule this course applies everywhere else, applied one more time to grading itself.

When it goes wrong

Symptom	Cause	Fix
Author grades their own capstone	No independent reviewer assigned	Route every submission through someone who wasn't in the room when it was built
High score despite a scope violation in the run history	Reviewer skipped checking the automatic-fail conditions first	Check auto-fail conditions before scoring anything else — they override the point total
Rubric passed, but the loop clearly wouldn't survive a real week of use	Scoring rewarded description over demonstration	Weight run-history evidence over design-document polish, per the point table above

Sources: this rubric operationalizes the T4 exit assessment named in learning-tracks.md ("design, justify, and govern a ... system against the rubric") and the minimum-safe checklist from CLAUDE.md rule 5, both grounded in Panaversity's Loop Engineering: A Crash Course (S1) and the maker≠checker principle of Step 11. Full attribution: resources/sources.md.