Full JSON and long-text benchmark report for website inclusion

Open index.html for the full report. It prominently separates the existing 10-task JSON cohort from the expanded 100-task long-text cohort; no pooled savings average is used. Includes TL;DR, cost and token summaries, answer correctness, decision reasons, targeted follow-ups and complete input ledgers.
explore.html contains every long-text source, answer, original score and source-aware judgment, plus the 0.25/0.15 follow-ups. json-explore.html contains all 10 structured JSON examples, all four representations, 120 existing answers and the new source-aware revalidation. Both explorers retain strike-through viewers. json-quality-evidence/ retains all 22 new judge calls and exact round-trip results; there were no new compression or answer-generation calls. Input links select the corresponding case.
Upload the directory together; no server, API credential, build step or remote data fetch is needed. Styles adapt to light/dark mode and narrow screens.
calculations.json retains exact metrics and evidence hashes; explore.json and CSV files retain the full underlying evidence. Original archives remain unchanged.
Costs are frozen API-equivalent scenarios dated 2026-11-08, not subscription invoices. No broad quality-preservation claim is made from the selected two-case follow-ups.
