AI-Run Modern LaTeX Manuscript Workflow and Replication Packet

This same-concept successor publishes the controlling July 28 native-diagram correction to the AI-run Modern LaTeX Manuscript Workflow. The public surface keeps the corrected seven-page A4 PDF and Markdown directly accessible, co-publishes the exact Claude cold-reverify method, the resource-efficiency incident note, and the diagram-fidelity correction, groups the current files into one source packet, and retains the July 6 addenda/scripts ZIP byte-identically. Top-level sessions own disjoint whole-expose ranges and retain responsibility for mathematical translation, transcription adjudication, diagram reconstruction, reference semantics, and visual PASS decisions. Subagents are limited to bounded mechanical support or preliminary drafting. Loop 1 advances canonical text and equations; Loop 2 handles native diagrams and exhaustive reference/release work without blocking disjoint Loop-1 production. For scan-controlled work, the controlling source image decides. Page mapping uses printed page, running header, and folio. The method calls for five overlapping 2400-dpi text bands per page, 300-dpi page context, about 5000-dpi default diagram comparison, targeted 9000-dpi crops for real ambiguities, and node-by-node and edge-by-edge review. Existing 600/1200-dpi evidence remains valid history and context; only 300-dpi-only approvals and independently identified material defects are reopened. Every diagram delivered in a new SGA3 reader or payload is native editable TeX. Raster crops remain private authority witnesses and are excluded from new public readers and payloads. Final successors bind disjoint top-level-session ownership and a lead-signed exact high-zoom review. Prior public checkpoints remain immutable history; material defects receive additive no-overwrite successors. The complete user-supplied OCR is a read-only locator and drafting witness. It must not be generated, rerun, re-extracted, or delegated. SGA1 and SGA2 are not blanket-retranscribed from images when their completed mathematical TeX transcription is already controlling. Source images are opened for genuine ambiguities, diagrams, or an explicit source-control question. The incident note records avoidable duplicate visual checks, repeated OCR/transcription activity, agent audit cascades, and repeated builds or manifests that did not advance the mathematical corpus. Its emissions figures are transparent scenario calculations, not metered OpenAI telemetry. Multi-ton coal-equivalent outcomes are conditional on the stated high-overhead, several-hundred-million-token assumptions; lifecycle, labor, infrastructure, and opportunity costs remain unquantified. This is a professional methodology and accountability publication, not a manuscript certification or new license grant. Documentation and exact release identities support production and preservation; they do not replace translation, transcription, diagram reconstruction, or public custody.