Repository navigation
docs: add sourced project slides and reproducible figures - #998
realmarcin wants to merge 5 commits into
Conversation
Build a 39-slide content package with seven reproducible figures, explicit evidence limits and source checks. Correct the geometry-library interpretation while preserving measured tables and numerical policies.
|
Fresh independent review of head00e6363 re-derived the seven figures and checked the narrative against committed sources. PRIMARY independently verified its findings:
Corrections and a fresh fix-diff review are pending. The separate follow-ups #964 and #987 remain outside this PR. No merge is authorized. |
Correct the logged Round7 sulfate inventory, remove an unsupported geometry bound, and preserve every Terms block in the rendering handoff. Retain exact scientific manifest and PNG checks while exposing cross-platform field, pixel and encoding differences for issue999.
|
Fresh fix-diff review and PRIMARY verification identified these remaining items:
For #999, downloaded Linux artifacts and PRIMARY's independent comparison show identical scientific metadata/source hashes. Two figures have identical RGB but different PNG serialization; five have tiny arrow-edge pixel differences. Exact checks remain required. A narrowly scoped arrow-raster and canonical-encoding repair is being prepared; no cross-platform success is claimed yet. |
Use canonical PNG encoding and stable annotation arrows while retaining exact figure checks. Guard diagnostic output paths, verify the encoder with independent decoders, and correct R5/R6 archive inventory and geometry-policy attribution.
|
Fresh-context fix-diff re-review completed on 855d3c8. Separate engineering and scientific-source reviewers verified the final committed blobs and found no new in-scope defects. The source reviewer reconfirmed the inherited generator/retained-evidence finding already tracked by #1003; it remains follow-up for the 1SAR project. PRIMARY independently verified actual side effects and ran the full locked hermetic gate, all eight benchmark+slides tests without skips, Ruff and whitespace checks successfully. Exact scientific manifest/source hashes are unchanged by PNG regeneration. Cross-platform CI remains the final acceptance check for #999; no merge is authorized. |
Avoid a redundant quadratic path in the remaining cross-platform arrow mismatch while preserving the exact image and scientific metadata checks. Record the observed CI evidence and the authoring-time verification boundary.
|
Accepted on reviewed head 2c8d8c6. Both validate jobs succeeded in run 42209075641. PRIMARY independently downloaded the Linux and macOS artifacts: all seven PNGs and figures.json match the committed files byte for byte. Local full hermetic gate, all eight explicit slide tests without skips, source recounts, decoder/guard tests and fresh-context fix reviews passed. No tolerance or skipped cross-platform check was introduced. All 36 in-scope findings are addressed; #964, #987 and #1003 remain follow-ups. The PR stays unmerged for the user. |
Include the pinned MAD_FLOOR definition in the slide and generated figure metadata. Validation: full hermetic gate with benchmark/slides extras; seven PNG hashes unchanged; fresh read-only review found no findings.
|
The threshold-discovery source citation correction is committed in cfe9b63. S09a and the generated Figure 7 metadata now explicitly cite the pinned MAD_FLOOR definition at scripts/bench_recover_leg.py:63. All seven PNG hashes are unchanged; numerical content and source pin are unchanged. PRIMARY ran the full locked offline gate with benchmark/slides extras successfully and verified the exact three-file diff. A fresh-context read-only review found no actionable findings. This addresses the additional #977 citation sibling. CI on the new head is pending; no merge is authorized. |
The earlier deck needed a clearer introduction and a clearer account of what each comparison establishes. This adds a Claude Slides content package with 39 slides, general-audience titles and technical subtitles. Each of six distribution figures has an introduction defining its terms and experiment, followed by a slide of sourced results and implications. XMechBay builds the visual deck from this package.
The narrative covers the assessment harness, incorporated tools, three benchmarked geometry configurations, the broader documented geometry taxonomy, threshold discovery, the extreme-resolution control set, retained results and the 1SAR evidence limits. Q3 records the user's taxonomy decision; unsupported project framing remains in questions.md. A new threshold-discovery triplet separates evidence and policy states across 41 scoped components, recounts the committed research records, and shows the Round 4/7/9 registration history. Scientific figures read pinned committed records at f3cfad3. A separate dated record documents filesystem presence on the M5 without implying benchmark coverage.
One generator reproduces seven committed PNGs and their SHA-256/source manifest. The change also corrects the bounded geometry-library interpretation and its provenance pins while preserving study tables and adopted numeric bands, and corrects the Round7 sulfate attribution and raw R5/R6 archive inventory from retained bytes. Historical evaluator observations keep their separate attribution, and missing Round7 coordinates remain log-only. Linux/macOS CI retains the core gate and adds the slides dependency check with full source history.
Package: data/coscientists/openscientist/slides_2026-11-06/.
Audit: ref/research/slides_content_package_review_2026-11-06.md.
Validation: the required locked benchmark-environment hermetic gate and the explicit locked benchmark+slides suite passed on Python 3.12. The latter exercises all eight tests without skips, including byte-for-byte regeneration of all seven PNGs and the manifest, PNG decoder/boundary checks, and diagnostic destination safety. Independent recounts match the ledger, document inventory and registered threshold history. Raw archive bytes independently verify the R5/R6 corrections. No PHENIX/CCP4, refinement or benchmark run was performed.
The original Linux failure was isolated to PNG encoding and annotation-arrow pixels; scientific metadata was identical. The repair uses a narrow canonical RGB8 encoder, disables antialiasing only on nine annotation arrows, and uses an explicit straight connector for the recovery-figure annotation. Exact comparison remains mandatory. Fresh independent engineering and scientific-source reviews verified the correction diff, followed by a fresh review of the final connector change; no unresolved in-scope finding remains. Both validate checks passed on reviewed head 2c8d8c6 in run 42209075641. PRIMARY downloaded both platform artifacts and verified all seven PNGs and the manifest byte for byte against that commit.
Addressed findings:
Closes #965.
Closes #966.
Closes #967.
Closes #968.
Closes #969.
Closes #970.
Closes #971.
Closes #972.
Closes #973.
Closes #974.
Closes #975.
Closes #976.
Closes #977.
Closes #978.
Closes #979.
Closes #980.
Closes #981.
Closes #982.
Closes #983.
Closes #984.
Closes #985.
Closes #986.
Closes #988.
Closes #989.
Closes #990.
Closes #991.
Closes #992.
Closes #993.
Closes #994.
Closes #995.
Closes #996.
Closes #997.
Closes #999.
Closes #1000.
Closes #1001.
Closes #1002.
All 36 in-scope findings are addressed in the reviewed head. This PR remains unmerged pending explicit approval.
Separate source follow-ups: #964, #987 and #1003 remain outside this PR. #1003 concerns future assessment-generator output and an explicit dated correction for retained evidence; historical raw records remain unchanged here.