[ethancjackson]

What

The BoC press-release HTML extractor grabbed the entire <main> element, so cached rationale artifacts carried trailing page furniture from modern (2021+) announcement pages:

• the content-type taxonomy line (Content Type(s): Press releases) — on all 139 cached releases;
• related-content link teasers appended after it;
• on 2025-01-29 specifically, a block of footnotes plus a bundled same-day operational notice (the deposit-rate/term-repo notice), which is why that file was an outlier at ~9,964 chars.

This is the text the reasoning-alignment evaluator (notebook 03) reads as the "official rationale," and it's the context seam available to the LLMP/agent predictors — so the cruft was participant-facing.

Fix

Truncate the extracted body at the earliest line-anchored footer marker, {Footnotes, Content Type(s)}:

• Content Type(s) is the universal boundary — drops the taxonomy line and the teasers that follow it on every page.
• Footnotes is the rare boundary (1/139) that also strips the 2025-01-29 bundled notice, which precedes the taxonomy marker.

Both markers sit alone on their own line and never occur in rationale prose, so the cut is safe and a near no-op on older (2009–2020) pages.

Verified on the regenerated cache (scripts/fetch_boc_press_releases.py --force):

• no footer marker remains in any of the 139 releases; Content Type(s) appears nowhere;
• 0 non-sentence-final endings (every release now closes on its rationale sentence);
• median 3844 → 3484 chars, max 9964 → 5590 (2025-01-29: 9964 → 4590).

Tests

Adds two fixture tests to test_press_releases.py:

• modern page (footnotes + bundled notice + taxonomy + teasers) → trimmed to the rationale;
• taxonomy-only page (older releases, no footnotes) → exercises the universal cut independently.

Full BoC suite: 44 passed. ruff clean.

Note: the report cache under data/reports/boc_press_releases/ is gitignored; reviewers regenerate it with uv run python scripts/fetch_boc_press_releases.py --force (HTTP only, no LLM/proxy calls).

🤖 Generated with Claude Code