What to do with the current gate bundles
The current bundled review instructions should survive, but their role should change.
Decision
Keep the current bundle names as stable user-facing entry points. Stop treating the bundle documents as the canonical place where checks live.
In practice:
frontmatter-review,prose-review,semantic-review, andcomplexity-reviewremain valid bundle names- their check lists move into individual gate files
- the bundle document becomes a wrapper containing shared framing, prerequisites, cautions, and output contract
This preserves the interface without preserving the wrong storage primitive.
What stays in a bundle wrapper
- what the bundle is for
- shared prerequisites
- bundle-level do-not rules
- output format
- optional execution guidance such as target budget or ordering
These are genuinely bundle-level concerns. They should not be duplicated across every gate.
What moves out of the bundle wrapper
- individual check definitions
- canonical bundle membership
- stale policy
- anything the selector needs to compute freshness
Those belong in gate files and derived indexes.
Why not delete the bundles outright
Deleting the bundles would create unnecessary UX churn.
- users already think in terms of named review passes
- execution may still want bundle-sized packets even after freshness becomes gate-local
- shared output and prompting constraints are real bundle-level concerns
So the right move is to thin them, not remove them.
Why not keep them canonical
Keeping the current bundle documents canonical would preserve the main problem:
- checks remain trapped inside monoliths
- selector logic keeps inferring semantics from bundle names
- the same check cannot cleanly appear in multiple bundles
- review-revise extracted gates have nowhere first-class to land
That would recreate the current coupling under a new label.
Specific recommendation for review-revise experiments
The workshop-only accessibility-review and sentence-review documents should not be promoted wholesale as new canonical bundles just because they surfaced useful checks.
Recommended treatment:
- harvest stable checks from them into individual gate files
- mark those gates
candidateoractivebased on confidence - then decide whether they belong inside existing bundles or justify a new named bundle
This matters because the review-revise experiment mixed two things:
- discovery of useful individual checks
- experimentation with review-pass packaging
We should preserve the first result without prematurely freezing the second.
Practical migration stance
Near term:
- keep current bundle files working
- decompose their checks into gates
- let selector freshness become gate-local behind the scenes
Later:
- make bundle files thin wrappers
- let execution regroup stale gates by bundle when useful
- retire duplicate monolithic check prose
That is the least disruptive path that still gets us to the right architecture.