[chaliy]

What changed

Markkit now runs WCXB and AgentDocBench through Mira with pinned inputs, the unchanged upstream scorers, release CLI conversion timings, and a committed quality baseline. PRs and pushes run 178 cases; nightly runs evaluate 2,754 cases. Reports and provenance are uploaded even when a gate fails.

The fixed development slice gates canonical aggregates and per-type/category/format scores, allowing at most 0.005 absolute regression. Full dev runs gate that identical subset and report the complete selection separately. Manual test evaluation enforces coverage/execution while keeping test quality report-only.

Why

The existing extraction study has ten handcrafted HTML fixtures. Public corpora add realistic web page diversity and native-document/PDF coverage, while the internal structural, QA and security tests remain in place.

Before / After

Before: no public-corpus run, public baseline, or scheduled public regression gate.
After: just eval-public-setup, just eval-public-prepare, just eval-public produce canonical quality and fresh-process conversion timings through Mira.

Executed release results, committed under crates/markkit-eval/results/:

• Fixed PR selection: 178/178 cases; zero regressions. WCXB F1 0.6771 default / 0.4999 main-content.
• Complete development selection: 2,754/2,754 cases; zero regressions. WCXB F1 0.7366 default / 0.6534 main-content.
• AgentDocBench: 0.7520 full category macro; supplementary no-OCR 0.8204. Three image-only PDFs remain zero in the full denominator.

The corpus audit found 139 identical HTML files shared by WCXB dev/test. Development runs exclude them, leaving 1,358 pages; the 511-page test split stays separate. Empty/absent full-text references retain upstream scoring behavior and are identified in reports.

Validation: just pre-pr passed; all 34 script tests passed in the scorer environment. Contract tests cover split leakage, verified-cache recovery, traversal, canonical weighting, missing/duplicate/N/A/invalid scores, regression policy, and failed runs replacing stale successful artifacts. Mira 0.5.0 and Python dependencies are checksum/hash pinned. Most added lines are immutable corpus locks, attribution and unchanged vendored SDK/scorers.

Risk

• Low: development tooling and CI only; no published library/binding dependency or extraction behavior change.
• External corpus/release availability can fail preparation; verified caching limits repeated downloads.
• CI gates quality, not latency across different machines. Timings use three sequential fresh processes after one warm-up and include startup.

Checklist

• Tests added or updated
• Backward compatibility considered
• Knowledge, usage docs, notices and Unreleased changelog updated