[v0.4.0] Research report PDF pipeline #47
Winnie
Loading…
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Prerequisite for the researcher-grade PDF report Robert requested on #39. Builds a professional PDF from the structured iteration log (#46) + game-rulings + issue history: title page, abstract, methodology (game model, Lisp AI architecture, bench harness with fixed seeds, 13 punishing presets), results (full hypothesis/experiment tables, per-preset mains table, per-specialist gauntlet table, before/after comparisons vs v0.3.0 via --bench-vs), key findings, accounting-validation notes (ACCT/CF divergence checks), future work (central bank, replay, churn economy). Must be presentable to a professional researcher. Deliverable: the generated PDF emailed to robert@machinetele.com as part of the #34 ship gate, and the generator checked into the repo for future releases.
Step 1/6 — dispatched, starting now. I'm the subagent on this issue.
Plan for the researcher-grade PDF report pipeline (#47):
47-reportbranched fromorigin/46-iter-login bankdevtools/report/(consumesiterations.jsonper the #46/#48 schema, plus git history + issue data via the Forgejo API; never fabricates numbers)%PDFmagic, page count, and extracted text47-report, open PR againstmaster, request review from robertismoI will comment again at every intermediate step. No bench runs will be executed.
Step 2/6 — worktree ready.
/root/bw-47-reporton branch47-report, branched fromorigin/46-iter-log@4bf460e([#46] Structured iteration log). Clean tree. This gives me the exactiterlog.goschema for the generator.Next: read
iterlog.go+bench.go(13 punishing presets) +bench-results.md, fetch issue #46/#47 narrative from the API, then build the generator undertools/report/. No bench runs, as promised.Step 3/6 — generator built.
tools/report/is drafted in the worktree:generate_report.py— CLI; consumes iterations.json (schema read straight offiterlog.goon 46-iter-log), reads git history viagit log, fetches issue data from the Forgejo API. Never fabricates: unknown values render as '—', schema validation results are reported in the PDF itself.bwpdf.py— vendored pure-stdlib PDF writer (no pip needed; only zlib etc.).templates/— 7 section templates (abstract, methodology, results, findings, accounting, future, provenance).testdata/iterations.sample.json— schema-valid synthetic fixture (2 records, 13 presets, 10 gauntlet specs, bench-vs rows); real git SHAs/timestamps, synthetic numbers, stamped via--note.README.md— deps, usage, schema, verification commands.Sections covered: title page, abstract, methodology (game model, Lisp AI architecture, bench harness + fixed seeds, the 13 punishing presets from bench.go), results (per-preset mains, specialist gauntlet, bench-vs before/after), key findings, accounting/validation notes, future work (central bank #15, replay, churn).
Next: end-to-end generation from the fixture + PDF verification (magic bytes, page count, text extraction).
Step 4/6 — end-to-end verification PASSED.
Generated three real PDFs from the fixture + repo and verified each with pypdf (skill's method) and gs rasterization:
iterations.sample.json, 2 synthetic records): 12 pages,%PDF-1.4magic ✓, 24,647 chars extracted, every expected token found (Preset/empire/mill/vault/adapttables, gauntlet, bench-vs, validation table,ACCT DIVERGENCE,Vickrey, SHAs5d760fc/be367deresolved from git history). Spot-checked fixture numbers: record-2 duel wins 13/47/13/13 all present in extracted text ✓. Visual check of rendered pages 1/3/6/12 via rasterization: title page, banners, ruled tables all clean.iterations.json=[]): 6 pages, reports "no records" instead of crashing ✓.--all-recordsrun: 15 pages, full detail for both records ✓.Found and fixed two writer bugs during verification (object references in the vendored PDF writer; UTF-16 metadata encoding). Everything runs on pure Python stdlib — no pip.
Next: ship files into the 47-report worktree, commit, push, open the PR.
Step 6/6 — done. PR open: #51 (branch
47-report->master, 12 files, +2,351 lines, review requested from robertismo). Master was not touched; no bench was run.How to run it
Requirements: python3 >= 3.8, standard library only — no pip. Verification recipe (pypdf, page count, text extraction) is in
tools/report/README.md.What the PDF contains
Title page, contents, abstract (headline stats computed from the log), methodology (game model, Lisp AI architecture, bench harness + fixed seeds, the 13 punishing presets), results (per-preset mains, specialist gauntlet, bench-vs before/after, iteration history), key findings, accounting/validation notes (schema validation the generator actually runs, reported as PASS/FAIL), future work (central bank #15, replay, churn economy), and a data-provenance appendix. Every number comes from iterations.json / git history / the Forgejo API — nothing is hand-typed; unknown values render as '—'.
Verified
12-page PDF from the synthetic fixture (
%PDF-1.4magic, 24,647 chars extracted via pypdf, fixture numbers spot-checked in extracted text, pages rasterized and inspected), graceful 6-page empty-log report, 15-page--all-recordsrun, TOC page numbers cross-checked. The generator consumes exactly the #46/#48 schema (iterlog.go).Merged: PR #51 (tools/report PDF generator) is on master. It consumes iterations.json + git history + live issue data, enforces no-fabrication, and produces the researcher-grade report. The actual report PDF gets generated and emailed as part of the #34 ship gate once the #39 mains work lands. Closing.