Calibration SubAgent Skill

Calibration SubAgent Skill

Goal

For the current TOC region, discover page-numbering regimes and an initial offset for each regime that has usable entries. Submit candidate offsets via calibration.submit. After submit, a deterministic completion pass runs the production tail-verify → binary-search → small-step recalibrate loop (using the same visual page confirmer as production). Only complete segments are usable for coarse structure; unrecognized pages are treated as no TOC.

Do not use

• Do not scan a fixed window after the TOC (the old “TOC end + N pages” probe).
• Do not invent physical pages you did not inspect or obtain from link.
• Do not track a total page-count budget. Limits are token budgets and max_rounds in the payload. A per-call page cap is only a batch-size limit.

Mandatory first step — partition regimes

Inspect every page_number label on the current TOC entries and partition them into page-numbering regimes (distinct numbering systems / label shapes: decimal digits, roman numerals, prefixed folio labels, etc.).

• Do not mix samples across regimes when computing an offset.
• Run the same initial-calibration procedure independently for each regime that has usable entries.

Phase 1 — Initial offset (your job via tools)

For each regime:

1. Select a small set of entries (prefer spread: early / middle / late when enough entries exist).
2. Candidate physical page:
   • If the entry has link.physical_page, use it as the primary candidate.
   • Otherwise derive a coarse physical candidate from the printed label and page_count, then confirm with vision.
3. Progressive inspect.pages for that title (start small, expand only if needed):
   • 1st call: inspect 1 candidate page only.