minxnnu-cloud Rework Budget
Rules for pipelines where an artifact is built, reviewed, rejected, and rebuilt. All of them were mined from a 52-round revision audit of a real production pipeline; over half of all rounds were repeats of a cause that had already appeared.
1. Minimal successor — close open defects only
A successor round closes the previous round's open defects and does nothing else. Everything unchanged is inherited by exact reference (path + hash), not regenerated, not re-audited. The audit's cleanest rounds were labeled "minimal successor: closed 2/2 open defects"; the messiest regenerated a 90-page artifact to change two pages.
Specs revise the same way: as a delta contract naming which clauses it supersedes — never a full rewrite. One full rejection triggered a 703KB spec rewrite; the delta that actually mattered fit in a page.
2. Same cause twice = escalate
When the same cause category rejects a round twice, stop adjusting at that level and go up:
- render tweaks repeating → the spec is wrong, fix the spec
- spec fixes repeating → the direction is wrong, re-lock direction with the user
- direction flapping → you're guessing; ask the user one interpretation question
A third attempt at the same level was almost always waste in the audit.
3. A human rejection becomes a machine gate — same round
Mechanical checklists catch what's wrong (fonts, breaks, colors). They cannot catch what's lacking — emptiness, density collapse, lost identity, underfill. Those arrive as human rejections. The only way the next deliverable doesn't pay the same human round: convert every human rejection into a quantitative gate (a number, a classifier, a grep) and append it to the checklist in the round where it happened.
4. Identity gates — machine-check at build time
Three defect families that each burned multiple rounds, all machine-detectable:
- Stale identity. Artifact metadata (PDF Document Info, XMP — including orphan objects via raw scan) must match the current course/revision; zero tokens from other courses or past revisions. A fix that cut the catalog reference but left the raw string got re-rejected one round later.
- Exact requirements. The user's exact copy and contract numbers (minimum font sizes, required notices) live in a machine-comparable list, checked at build. Four audit findings in one round were all trivially automatable.
- Single canonical mapping. When two frozen inputs can disagree (a TOC's page numbers vs. rendered folios), name one canonical source in the binding input and machine-compare the other against it. Two "frozen" canons that silently disagreed cost an entire 90-page candidate.
5. Lock only what deserves locking
Hash-lock tables belong on frozen canonical inputs — not on every document. In the audited workspace the receipt ritual had spread to 130 of 148 documents (2.79M characters total). When everything is locked, locks stop being signal.