A lease amendment chain answers as-of questions

[WaylandYang]

Adds the lease-chain bench used to find and measure #679 and #680. It is scripts only; no filings are committed, the same as the NVIDIA corpus.

The recall bench (recall.mjs) asks whether a fact reached the graph. This one asks whether, after a contract is rewritten by a chain of amendments, its terms answer correctly at a moment. The Blackbaud HQ lease's Phase 2 exercise deadline was extended five times in 2020 before the tenth amendment tied it to a trigger date. The property was sold to a Blackbaud affiliate that became the landlord, and the price was cut after inspection. There are 18 questions about those values on particular dates.

What is added, under scripts/bench/corpora/contracts/

• edgar-manifest.json lists 7 contract chains and 62 EDGAR filings. For each it records CIK, accession, filing date, role, and a title read from the document's opening.
• fetch.py fetches them into edgar/, as HTML plus plain text. It uses curl for the local proxy and needs SEC_USER_AGENT. The redacted CSG/Comcast chain is fetched only with --with-redacted.
• lease_bench.py setup creates a fresh base and turns off the switches that edit the graph after extraction, as recall.mjs does. It declares a generic lease and property-sale ontology, with single-valued terms marked functional, and ingests the 13 amendments (the 145-chunk base lease only with --with-base). It logs in like recall.mjs: BENCH_BASE, and BENCH_TOKEN or BENCH_EMAIL/BENCH_PASSWORD.
• lease_bench.py score answers the questions the way the product reads. It takes each lease and purchase-agreement entity's facts and keeps those whose server-computed holds_from/holds_to contains the moment. A question is right when the values holding then are exactly the true one.
• README.md covers how to run it, the truth table, and the 2026-09-13 runs.
• .gitignore ignores the fetched edgar/, cuad/ and the per-run lease-kb-*.json.

A run is one fresh base and takes about 10 minutes. Runs vary by about 3 questions, so a build needs at least two.

🤖 Generated with Claude Code