A review layer that survives the next draft.
Project Meadow keeps decisions attached to the claims that caused them—even when an AI-generated document changes shape.
Extension test bench
Three revisions, three anchoring cases
- Baseline: add comments to highlighted test phrases below.
- Ordinary edit: verifies quote recovery, duplicate context, and aliases.
- Destructive edit: removes one target so its thread should become detached.
01
Executive summary
Review should be part of generation, not a cleanup step.
Teams now review generated PRDs in chat threads, screenshots, and copied excerpts. The discussion is useful, but the document cannot carry those decisions into its next generation. Meadow introduces an independent annotation layer with durable text anchors and explicit resolution events.
The outcome is a review trail that remains legible to people and deterministic to generators. Open discussion stays open; accepted requests can change the next version; refused requests remain visible without changing the source.
02
Problem
The feedback exists. Its coordinates do not.
A reviewer can identify a risky sentence precisely in the current draft, but a regeneration may insert a paragraph, rename a heading, or rebuild the entire DOM. URL-only comments survive the deployment while losing the passage they were meant to discuss.
“A useful review system must preserve intent without pretending every old coordinate is still correct.”
Silent loss
Comments disappear when generated markup changes.
False recovery
A thread reattaches to similar but unrelated text.
Implied consent
A generator treats silence as permission to edit.
03
Decision policy
Conversation is context. A decision is an event.
A decision is valid only when an editor explicitly accepts or refuses the request. Conversation text can explain what should happen, but it never changes the effective decision by itself.
In the review workflow, the review layer must never infer consent from silence. An unanswered thread remains open and the next generator carries it forward without applying it.
04
Scope
Start with text, threads, and trust boundaries.
In the first release
- Select document text and start a thread.
- Reply chronologically with author and time.
- Isolate review data by Workspace.
- Accept, refuse, or reopen through explicit actions.
- Recover anchors after ordinary document edits.
Later
- Pixel, canvas, video, and PDF regions
- Live cursors and reviewer presence
- Reactions, mentions, and notifications
- Automated fuzzy reattachment approval
- Multiple acting identities per user
05
Architecture
Keep page interaction local and authorization server-side.
The page never receives credentials, membership details, or another Workspace’s annotations. Selected quotes are the only source content stored by default; full pages are not uploaded.
06
Risks
Honest failure is safer than confident misattachment.
Native product overlap
Host platforms may add richer annotations. Cross-platform Workspaces and audit semantics must remain independently useful.
Generated DOM churn
When both the stable section identifier and the selected quote disappear, the thread must move to the detached tray instead of attaching to nearby prose.
Permission fatigue
Use temporary active-tab access by default and request persistent host access only for verified artifact frames.
In the generation pipeline, the review layer must never infer consent from silence. This duplicate phrase has different surrounding context so quote disambiguation can be tested.
07
Rollout
Prove the anchor, then close the collaborative loop.
- M1AnchoringBuilt
Select, comment, reply, reload, recover.
- M2WorkspacesBuilt
Google sign-in, invitations, authorization, realtime sync.
- M3Generation loopBuilt
Stable IDs, OAuth context, conversations, receipts.
Working principle
Preserve the discussion. Require the decision. Never fake the coordinate.