[Gartenberg sub C] Design: block-level prose provenance (model-drafted / human-authored / human-edited) so a disclosure statement can be evidence-backed

Sub-issue of [META] Gartenberg et al. 2026 (#829)

Anchor

• Editorial §5.1: the authors used AI for code, outlining, phrasing, and comparison, revised extensively by hand, and disclose the classifier score of their own text (8.8%). §5.4.1: detection should inform editorial triage, not gatekeep; editors want to "understand how a manuscript was produced and evaluate it in context." §3.4: outcomes are indistinguishable below ~30% AI share and deteriorate above it.
• ARS files: academic-paper/references/disclosure_mode_protocol.md (venue statements rendered from use records such as EDITED / FORMATTED_CITATIONS), shared/contracts/patch/block_manifest.schema.json and scripts/ars_anchorize_draft.py (#390: every block has a stable id and hash), scripts/ars_apply_revision_patch.py (apply report records which blocks changed), #106 (disclosure schema discovery, closed), docs/DATA_FLOWS.md.

Problem

ARS can say which venue policy applies and which categories of use occurred, but it cannot say, for a given manuscript, which passages the model drafted, which the author wrote, and which the author rewrote after drafting. That is the production-mode information the editorial says editors want, and the information that would let an author make a human-first claim with evidence instead of assertion. Without it, a disclosure statement is narrative.

Proposed change (design issue, needs-design)

1. Extend the block manifest with an optional per-block origin record: model_drafted, human_authored, human_edited_after_model (with a normalized edit-distance share between the model's last emitted text and the current block), plus the pipeline event that set it. Set deterministically by the anchorize step (initial origin) and the apply script (edits through the controlled path); manual edits outside the controlled path are detected by hash mismatch at the next anchorize and recorded as human_edited_after_model with unknown share.
2. A summary in the Process Record (Stage 6) and an optional line in disclosure mode: "N of M blocks human-edited after drafting; K blocks human-authored; word-share X%". Numbers come from the manifest; the renderer never estimates.
3. Resolve in the design: how this interacts with the #390 patch chain and freshness rule; whether origin belongs in the manifest or in a sidecar; what to record for blocks that predate the field (unknown, never backfilled, same rule as read_scope).

Acceptance

• Design doc under docs/design/ with the schema delta, the event table, and the unknown-origin rule
• Decision recorded on manifest-vs-sidecar placement
• Explicit statement that this is self-reported provenance from pipeline events, not a classifier score, and that it cannot see edits made outside the pipeline except as "changed, share unknown"

Non-goals

• No AI-likelihood classifier. The record says what the pipeline did, not what the text "looks like".
• No threshold that changes any gate. Whether a share is acceptable is the venue's and the author's call.
• Not a substitute for the venue-specific disclosure wording that disclosure mode already produces.