Close out the PDF study with a one-page summary - #740
Merged
Merged
Conversation
Records what the PDF research demonstrated, what it observed on a bounded population, and what it does not support, and retires the external-validity investigation as inconclusive. This is a replacement for the closeout proposed in #728, built from that PR's base. It adds no scored outputs, register entries, blind identifiers or individual judgments: those cannot be verified from public material and some pair control answers with identifiers. It does not amend the frozen rules; the §5.6 versus §7 disposition conflict is named and left open. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XmeDh5BVXMBkwZ7zCX1oca
This was referenced Sep 30, 2026
The next-task paragraph said the XML path reaches the module-level pypdfium2 import only through bill_tree.py. That was true when the delivery study checked it, but the canonical JSON formatter now imports parsers/pdf_anchors.py at module level too, so there are at least two routes. Every route ends at pdf_text.py, so the recommended task, making that import lazy, is unchanged; only the explanation was wrong. Links the task to #751. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01XmeDh5BVXMBkwZ7zCX1oca
…he closeout verify_parity.py --mutate returns 0 when it detects the injected mismatch and 1 when it does not, so the reproduce step now says exit 0. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
This was referenced Oct 4, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes out the PDF research with a one-page summary and retires the external-validity investigation as inconclusive. Supersedes #728 (the earlier closeout, which also published scored outputs and register entries).
What changes
docs/research/pdf-backend-bakeoff/CLOSEOUT.md: what the PDF work demonstrated, what it observed on a bounded population, what it does not support, the reusable code, and one next product task.validation/README.md, the directory's "start here" document.Documentation only. No code, probes, frozen inputs, thresholds, methodology or register entries change.
Why this replaces #728
#728 added
metrics.jsonand register entries A57 to A62.4 alongside its closeout. Those carry per-control answers from both adjudication routes keyed by blind ID, which join to the expected answers already public incontrol_fixtures.json, and they cannot be verified from public material because the oracle provenance is private. This version links to the existing public evidence instead and publishes no scores, blind identifiers or individual judgments.It does not undo what the #728 branch already exposed, and it says so. All 20 adjudication controls are treated as retired.
Things a reviewer should check
staffer-delivery/README.md,RESULTS.mdand its audit,validation/). Nothing was re-measured for this PR; the Pyodide parity result cites its 2026-08-11 re-run with a negative control.Next product task named in the document
Make the module-level
pypdfium2import insrc/deltatrack/parsers/pdf_text.pylazy, so the engine loads under Pyodide without a stub. The XML path never calls PDFium, but reaches that import through module-level imports ofparsers/pdf_anchors.py, at least frombill_tree.pyand the canonical JSON formatter. Making the import lazy inpdf_text.pycovers every route. Tracked in #751.🤖 Generated with Claude Code
https://claude.ai/code/session_01XmeDh5BVXMBkwZ7zCX1oca