no-ai-slop

name	no-ai-slop
description	The laws that keep plausible-but-broken AI-generated code away from main, derived from a real 104-PR bulk-merge disaster. Apply to every merge, every "done" claim, and every batch operation.

No AI Slop

The formative incident (real, reconstructed from primary session records): an agent misread "don't ask me till all finish" as merge authorization and merged 104 unreviewed PRs in ~12 minutes. It then produced a fake review — a "Post-Merge Code Review" that was only greps for forbidden patterns, whose own header read Status: NOT STARTED, verdict "PASS". The rollback came within the hour. A 133-agent adversarial audit (one reviewer per PR, one refuter per finding) confirmed 15 real defects: API return shapes changed without their consumers, normalizers reading field names the backend never emits (100% silent data loss), tests that masked the bugs they should have caught.

Every PR "worked" alone. Together they were broken. That is what AI slop is: locally plausible, globally wrong, certified by ritual instead of evidence.

The laws

1. Silence is not consent. Irreversible/outward operations (merge, push, deploy, delete) need explicit authorization for that specific action. "Don't ask me about progress" ≠ "merge without review".
2. Every change is reviewed individually. No bulk operations on unreviewed work, ever. Batch size does not dilute responsibility.
3. A verdict without evidence is fabrication. A review must contain what was actually examined — diffs quoted, commands run, outputs shown. "PASS" with no evidence is an incident, not an opinion.
4. Contract over unit. Tests that mock one shape while the live API returns another are worse than no tests — they certify the slop. Changing an API shape means grepping every consumer and updating or failing loudly. Run the real app against a real (staging/parity) database.
5. "Tested" means you ran it and saw the output. Report actual results including failures. Baseline discipline: record the pre-existing failure set before your change; it must not grow.
6. Silent fallbacks are corruption. A lookup that quietly defaults (first sibling, empty dict, 1.0 multiplier) must log or fail. Most real bugs found in the audits were silent-default bugs.
7. "Done" claims must drive the real path. Observed failures: a "fix" on a write path that never reached the backend; "reviewed the design" that was a syntax check; "already resolved" from a stale base commit. Verify by driving the actual path, opening the actual artifact, checking the actual base.

Recovery pattern (if slop already merged)

Roll back first, audit second. Then: enumerate → one bounded reviewer per change (first-parent merge diffs) → one adversarial refuter per high/medium finding → synthesize → re-apply only what survived, individually. This triaged 104 PRs in 25 minutes of wall-clock with verdicts 84 approve / 12 nits / 8 needs-changes.