Repository navigation
Refresh checksum inventory and repair Wave 4 hash key (supersedes #15) - #27
Conversation
Keeps main's 39 paths in order and hashes each from the merged bytes. Five lines change: verify_all.py, wave5_unified_forces_stochastic_qm.py, formal/verify_exact_rational_proofs.py, formal/GravityClaims.lean and formal/Wave12QuasarCollapse.lean. The Wave 4 certificate line already matched the repaired certificate this branch carries, and the README line already matched main's README. sha256sum -c SHA256SUMS now passes all 39 entries. CHECKSUM_PROVENANCE.md records the cause of each of the six failures main had. Signed-off-by: Orda <342776047+orda-aska@users.noreply.github.com>
ahrazzle
left a comment
There was a problem hiding this comment.
The manifest repair is right. The merge of ca55815cd2ca6f43356cc7ad5a7148fa83cec030 and b7550b10a9f70c546e5004e98fff3e6065717825 matches git once the single SHA256SUMS conflict is taken from main, and 02ad9a49 then regenerates that manifest in place. All 39 entries match the merged files. The Wave 4 certificate change is only the key name sha_26_canonical to sha256_canonical, with the same value. What should not land as written is CHECKSUM_PROVENANCE.md: the historical section names the wrong commits, miscounts the f8030d6 mismatches, and says pull requests 10, 12, 13, and 14 are still absent from main.
There are no published reviews and no inline threads on #27 or #15. The only conversation comment on #15 is the pointer issuecomment-6884933474. The defects #15 described in its body are fixed in this tree. The false history in the provenance note #15 added is not.
What the diff is
Against main b7550b10 the diff is three files: CHECKSUM_PROVENANCE.md added, SHA256SUMS with five lines changed, and results/wave4_unified_certificates.json with the key spelling changed. 5734446a keeps 0bf72bf and ca55815c as ancestors. Its tree is 99cc5d36747bbb2c1f3fc0e8151213106b5dc71a, the same tree git produces by merging the two parents and keeping main's SHA256SUMS. The only conflict is that file.
The regenerated manifest keeps main's 39 paths in the same order. SHA256SUMS and CHECKSUM_PROVENANCE.md stay out of the list. sha256sum -c SHA256SUMS --quiet exits 0, and the verbose run is 39 OK and 0 FAILED. The five changed lines are verify_all.py, wave5_unified_forces_stochastic_qm.py, formal/verify_exact_rational_proofs.py, formal/GravityClaims.lean, and formal/Wave12QuasarCollapse.lean. Each new digest equals the file in this tree.
The Wave 4 line is unchanged at 7bb2326b36aa… and equals the file committed in 0bf72bf. On main the same line is already that digest, while the file on main still hashes to 824d1a5d68f1… and still contains sha_26_canonical. Replacing that key with sha256_canonical and leaving the value 1a7f2ff0182b66d1a081e59e750b493afa6beaab0d976c5af2c6c72652d4452d is the entire JSON diff. That digest was already stored in SHA256SUMS at 8e9af4b, before any committed copy of the file had those bytes. 0bf72bf is the first commit whose file matches it.
The README line is unchanged and is efdc6462ab7e6de35ea4e8e0d24c55b26fd5edb06f91e8102ba695bbc7a4e202, which matches README.md here and on main. The PR body quotes that digest. 0b347aed does not occur in git log --all -S 0b347aed.
On main, sha256sum -c SHA256SUMS --quiet exits 1 with exactly six failures: the five files above plus results/wave4_unified_certificates.json.
History of the three phantom digests
Across every blob reachable from all 152 commits on branches and pull-request heads (386 blobs hashed), these three manifest digests from main never appear:
verify_all.pymanifest1e907f6af0b1…. The only committed version is49bd79489023…, introduced byfab0101.7f388dbdoes not change the file.formal/verify_exact_rational_proofs.pymanifest606431af2e13…. The only version isd0acfe8ea3ab…, fromd286264.formal/GravityClaims.leanmanifestdb6e4a202d35…. The only version isc2e635f3af90…, from86d1ee6.
The other two refreshed lines do have an older committed version, as the new section says. 290a9bbb1ddd… is wave5_unified_forces_stochastic_qm.py at 102a1ae. 942a9f2 changes it to effc8ef25d44…. 643bb2ed80cc… is formal/Wave12QuasarCollapse.lean at ecc4823. c326de3 changes it to 40b210c201e7…. That chirp-mass commit is on main and was not on ca55815c. The merge brings it in, and the new manifest hashes the merged file.
Checks
lake build with Lean 4.35.0 prints Build completed successfully (15 jobs). It also warns that hPos in formal/Wave10JeansDerivation.lean is unused. That file is not in this diff.
An axiom scan of the 62 Foundation0 declarations reports axioms used #[propext, Classical.choice, Quot.sound] and extras 0.
--check exits 0 for Waves 3, 4, 5, 6, 7, 9, 10, 11, and 12, and for verify_all.py and formal/verify_exact_rational_proofs.py. Wave 7 was rerun here and passed. Each wave checker prints CHECK OK (byte-identical). verify_all.py and formal/verify_exact_rational_proofs.py rewrite their certificate files even under --check. git diff --exit-code on those certificates is 0, so the rewrite matched the committed bytes. git status --porcelain shows no tracked changes.
On main, python3 wave4_unified_field_solver.py --check exits 1 with AssertionError: Canonical SHA-256 mismatch on disk.
python3 research/update-d144448-review/implementation/check_docs.py . --output-dir /tmp/gravity-d144448-checks prints Unrelated manifest mismatch count: 0. test_wave11_12_corrections.py runs 7 tests, OK. acceleration-evolution runs 13 tests, OK.
At this head, GitHub Actions regenerate (run 43646304292) and reproduce (run 132604769747's suite, run 43646304304) both completed with success. regenerate is the convergence-probe job: 6 unit tests OK, then PASS all nine probes. reproduce is the value-assessment job: 8 unit tests OK, then PASS four probes and six immutable source hashes. Neither job runs sha256sum -c, a wave --check, or lake build. The legacy combined status is pending with no statuses. mergeable_state is blocked with no reviews.
#15
#15's body said main at d144448 had 11 manifest mismatches and a misspelled Wave 4 key. Recomputing d144448 gives those 11 paths: verify_all.py, the Wave 5 generator, the exact rational verifier, GravityClaims.lean, the Wave 5/6/7 Lean files, the Wave 4 certificate, README.md, CLAIMS.md, and CHANGELOG.md. On this head every current manifest entry matches, and Wave 4 --check passes. Those defects are resolved.
The provenance note is not. Relative to ca55815c, #27 only appends 17 lines. The earlier table is unchanged, and its digest columns do match 364fa3e and the f8030d6 manifest (0 of 12 hash cells wrong). The commit labels and two sentences do not.
7f388db changes only METHODS_AND_PLAYBOOK.md (44 insertions). At 364fa3e the last commit for each of the eight rows labeled 7f388db is something else: fab0101, 942a9f2, d286264, 86d1ee6, 306e1e7, 29b6310, b2c93b0, and 2f0f059. The new section already names fab0101, d286264, and 86d1ee6 for three of those files, so the same document contradicts itself.
At f8030d6, nine files disagree with the manifest, not eight. CLAIMS.md is already 92ea1300… against manifest c459b238…, and it changes again at 364fa3e. Lakefile, README, and the notebook do agree at f8030d6 and change afterward. The sentence that splits the twelve rows into eight prior mismatches and four later edits is wrong for CLAIMS.md.
The sentence that pull requests 10, 12, 13, and 14 have not been composed on main was true of d144448. Their tips are ancestors of b7550b10, including merge 71664e1 for #10 and b9325fd for #14. This branch merges that main, and the sentence is still present tense.
The digest columns, the 30-to-36 addition of the six Wave 11/12 paths, the nine mismatches at e5d7ee8, and the statement that ca55815c differs from 0bf72bf only in documentation and SHA256SUMS all check out. ca55815c versus 0bf72bf is CHANGELOG.md, CHECKSUM_PROVENANCE.md, METHODS_AND_PLAYBOOK.md, RESEARCH_NOTEBOOK_AND_REASONING.md, and SHA256SUMS.
index.html still says 36/36 SHA256SUMS and 36 SHA-256 verified files. Main already has 39 entries, so that copy was stale before this PR. research/update-d144448-review/implementation/tree-hashes.json records SHA256SUMS content hash 62c6e0cf…, which is the blob in 2fd7c77, b27d247, and af166bb, not main (cdf7389a…) and not this head (5bd457df…). check_docs.py rewrites that packet. It does not compare the committed copy. The value-assessment input manifest still pins Wave 12 at 70a1204 to 643bb2ed…, and the copied input file still has that digest. The root file is the later c326de3 bytes. Those consumers are outside this diff.
Findings
-
The merge commit matches git's merge with main's
SHA256SUMS.git merge-tree --write-tree --name-only ca55815c b7550b10exits 1 and names onlySHA256SUMS. Aftergit read-tree -m d144448 ca55815c b7550b10andgit update-indexof main's blob8ad68320c2fe05e455cea29e15f7a49b52d316e8,git write-treeprints99cc5d36747bbb2c1f3fc0e8151213106b5dc71a, which is5734446a^{tree}. -
The regenerated manifest matches the tree.
sha256sum -c SHA256SUMS --quietexits 0. The verbose tally iscount_ok:39andcount_fail:0. A line-by-line recompute reportsOK 39/39andchanged 5, and the same path order as main. -
The three phantom digests are absent from history.
git rev-list --all --objectsplus SHA-256 of every blob reportsversions=1 wanted_present=Falsefor each of the three files, and a repo-wide scan reportshits 0for1e907f6af0b1,606431af2e13, anddb6e4a202d35. -
Wave 4 fails on main and passes here. On
b7550b10,python3 wave4_unified_field_solver.py --checkexits 1 withCanonical SHA-256 mismatch on disk. On this head the same command exits 0 and printsCHECK OK (byte-identical): wave4_unified_certificates.json sha256_canonical=1a7f2ff0182b66d1a081e59e750b493afa6beaab0d976c5af2c6c72652d4452d. -
The provenance last-commit column is false.
git show --stat 7f388dbisMETHODS_AND_PLAYBOOK.md | 44 ++++++++++++++++++++++++++++++++++++++++++++and nothing else.git log -1 364fa3e -- <path>names the eight commits listed above, not7f388db. -
CLAIMS.mdalready disagreed atf8030d6. The manifest entry isc459b23829444a125fd0b682ab01dc34441088566292b9dad13566ee3b29b6dfand the file is92ea13005f4ff98b2737b632ac54084b87305f79ac87f0d6fa8f19124c27071d. Nine of the twelve table files disagree there. -
Pull requests 10, 12, 13, and 14 are in main.
git merge-base --is-ancestorof each PR tip againstb7550b10succeeds, and it fails againstd144448. -
Lean, axioms, docs, and tests pass on this head.
lake buildends withBuild completed successfully (15 jobs).The axiom command printsscanned 62,axioms used #[propext, Classical.choice, Quot.sound],extras 0.check_docs.pyprintsUnrelated manifest mismatch count: 0. Wave 7--checkexits 0. Unit tests: 7 OK and 13 OK. -
CI at this SHA is green for the two probe jobs only. Check runs
regenerateandreproducehaveconclusion: success. The regenerate log ends withRan 6 tests in 0.001s/OKandPASS all nine probes. The reproduce log ends withRan 8 tests in 0.004s/OKandPASS four probes and six immutable source hashes.
Fix list
- In the inherited-mismatch table, replace each
7f388dblast-commit cell with the commit that actually last changed that file as of364fa3e:fab0101,942a9f2,d286264,86d1ee6,306e1e7,29b6310,b2c93b0, and2f0f059. - Rewrite the sentence that those eight entries last changed at
7f388db. Their bytes do disagree with thef8030d6manifest. That commit did not change them. - Correct the eight-versus-four split. At
f8030d6,CLAIMS.mdalready disagrees with the manifest, so nine files disagree. It also changes later at364fa3e. - Qualify or remove "These independent changes have not been composed on main." Pull requests 10, 12, 13, and 14 are in
b7550b10, which this branch merges. They were not ind144448.
Addresses proof review 5907449532 on #27. - The eight last-commit cells that cited 7f388db now name the commit that last changed each file as of 364fa3e (git log -1 364fa3e -- <path>): fab0101, 942a9f2, d286264, 86d1ee6, 306e1e7, 29b6310, b2c93b0 and 2f0f059. 7f388db only touches METHODS_AND_PLAYBOOK.md. - At f8030d6, nine of the twelve table files disagree with its manifest: those eight plus CLAIMS.md (92ea1300 against c459b238). lakefile.toml, README.md and RESEARCH_NOTEBOOK_AND_REASONING.md match there and change later. - PRs 10, 12, 13 and 14 were absent from d144448 and are in main b7550b1; the note now says so. - index.html now cites 39/39 SHA256SUMS and 39 verified files, the live manifest count. Signed-off-by: Orda <342776047+orda-aska@users.noreply.github.com>
ahrazzle
left a comment
There was a problem hiding this comment.
The parent is 02ad9a49cd2b97ad6e2110c9725d2b71daebb518. This commit answers review 5907449532. The delta is CHECKSUM_PROVENANCE.md and index.html only. The four requested corrections are in the note, and each one matches git. index.html now says 39/39, and the root manifest still has 39 lines that all check. tree-hashes.json is unchanged. Nothing in the tree compares that committed file to the working tree. check_docs still reports zero mismatches. The two named Actions runs are green on this exact head.
The same stat line in index.html still says 6 stock Lean 4 modules. lakefile.toml has 7 [[lean_lib]] entries. That 6 was already on the page before this commit, and this commit did not claim to change it.
The last-commit column is headed "as of 364fa3e". git log -1 --format='%h %s' 364fa3e -- <path> matches every quoted subject, including fab0101, 942a9f2, d286264, 86d1ee6, 306e1e7, 29b6310, b2c93b0, and 2f0f059. The comparison printed subject mismatches 0.
Those eight files have the same bytes at f8030d6 and at 364fa3e. For each path, the script printed identical=True and commits_between=0 on git rev-list --count f8030d6..364fa3e -- <path>. The prefixes agree, for example verify_all.py is 49bd79489023 at both commits and the Wave 4 certificate is 824d1a5d68f1 at both.
At f8030d6, nine of the twelve table files disagree with that commit's manifest and three match. The nine are the eight above plus CLAIMS.md. The file hash is 92ea13005f4ff98b2737b632ac54084b87305f79ac87f0d6fa8f19124c27071d and the manifest entry is c459b23829444a125fd0b682ab01dc34441088566292b9dad13566ee3b29b6df. lakefile.toml, README.md, and RESEARCH_NOTEBOOK_AND_REASONING.md match. The script printed disagree 9 and match 3.
git log --oneline f8030d6..364fa3e -- CLAIMS.md prints exactly two commits: c36bd0f Wave 11: add audited claims C-28..C-30 to CLAIMS.md and 364fa3e Wave 12: add audited claims C-31..C-33 to CLAIMS.md. c36bd0f is an ancestor of 364fa3e, and f8030d6 is an ancestor of c36bd0f. git show --stat c36bd0f is CLAIMS.md | 13 +++++++++++++. The parent blob starts 92ea13005f4ff98b and the c36bd0f blob starts 10e673891b53d43a. 364fa3e changes it again.
git merge-base --is-ancestor of the tips of pull requests 10, 12, 13, and 14 fails against d144448 and succeeds against main b7550b10. The tips are 57fece7db800, 71e33aab20f1, cce0284d03ef, and b473fc9c3037.
sha256sum -c SHA256SUMS --quiet at this head exits 0. The verbose tally is ok 39 fail 0. index.html is not one of those 39 paths, so this commit does not change SHA256SUMS. The hero text is 39/39 SHA256SUMS, and the scope line says 39 SHA-256 verified files. No other manifest count of 36 remains in index.html. The previous index.html digest 18b6b7c88063… is not stored anywhere in the tree (git grep exit 1).
git show of research/update-d144448-review/implementation/tree-hashes.json is identical at 02ad9a49 and at this head. The only code reference is check_docs.py writing a new packet to its output directory. It does not read the committed file. The review README only says that the committed packet records the prepared tree.
python3 research/update-d144448-review/implementation/check_docs.py . --output-dir /tmp/gravity-d144448-checks-b prints Documentation, checksum table, changed hashes and ten frozen files pass. and Unrelated manifest mismatch count: 0, then exits 0. The worktree stays clean.
Check runs on 3ede78bdc27b6ebcb00840688b670cbe4b5d4cb0 are both completed / success. regenerate is run 43647971814. reproduce is run 43648008707. Both are workflow_dispatch jobs for the convergence probes and the value-assessment probes, and both record this head SHA.
Supersedes #15. This branch starts from #15's head
ca55815cby ahrazzle, so its two commits (the Wave 4 key repair and the checksum refresh) stay in history. It then merges mainb7550b10(after #21 and #23 through #26) and regenerates the rootSHA256SUMSthe way #15 does: hash every listed file from the merged bytes, excludingSHA256SUMSandCHECKSUM_PROVENANCE.md.Head
3ede78bd, parent02ad9a49.5734446ais the merge (parentsca55815c,b7550b10) and takes main's 39-entry manifest.02ad9a49regenerates it and adds a provenance section.3ede78bdcorrectsCHECKSUM_PROVENANCE.mdand the manifest count inindex.htmlper proof review 5907449532; the manifest and README are unchanged.Changes in 3ede78b
7f388db(which only touchesMETHODS_AND_PLAYBOOK.md) now name the commit that last changed each file as of364fa3e, checked row by row withgit log -1 364fa3e -- <path>:verify_all.pyfab0101,wave5_unified_forces_stochastic_qm.py942a9f2,formal/verify_exact_rational_proofs.pyd286264,formal/GravityClaims.lean86d1ee6,formal/Wave5StochasticForces.lean306e1e7,formal/Wave6Convergence.lean29b6310,formal/Wave7ThresholdAudit.leanb2c93b0,results/wave4_unified_certificates.json2f0f059. The column header now says "as of364fa3e".f8030d6: nine of the twelve table files disagree with its manifest, the eight above plusCLAIMS.md(92ea1300againstc459b238).lakefile.toml,README.mdandRESEARCH_NOTEBOOK_AND_REASONING.mdmatch there and change later. The note now says that in place of the eight-versus-four split.d144448and are in mainb7550b10.index.htmlnow says39/39 SHA256SUMSand39 SHA-256 verified files, the live manifest count. It said 36 on main.research/update-d144448-review/implementation/tree-hashes.jsonis left as is. It is the hash packet of the prepared review tree (the packet README says it "records the prepared tree"),check_docs.pywrites a fresh copy to its output directory, and nothing compares the committed copy. ItsSHA256SUMSvalue62c6e0cf...is the Correct Wave 11 and 12 numerical benchmarks and physical claims (supersedes #13) #21-era snapshot, so it is a frozen record.The six lines that failed on main
verify_all.py1e907f6af0b149bd79489023fab0101.wave5_unified_forces_stochastic_qm.py290a9bbb1dddeffc8ef25d44102a1aeversion;942a9f2changed the file without updating the manifest.formal/verify_exact_rational_proofs.py606431af2e13d0acfe8ea3abd286264.formal/GravityClaims.leandb6e4a202d35c2e635f3af9086d1ee6.formal/Wave12QuasarCollapse.lean643bb2ed80cc40b210c201e7ecc4823version; the chirp-mass fixc326de3changed the file without updating the manifest.results/wave4_unified_certificates.json7bb2326b36aa7bb2326b36aasha_26_canonicalkey (824d1a5d68f1). The manifest digest is the repaired file from0bf72bf, which this branch carries, so this line is unchanged.The README line is
efdc6462ab7e6de35ea4e8e0d24c55b26fd5edb06f91e8102ba695bbc7a4e202. It already matched main's README, which this merge leaves unchanged.Checks on this head
sha256sum -c SHA256SUMS: 39 of 39 OK, zero failures (--quietexit 0).--checkon Waves 3, 4, 5, 6, 9, 10, 11 and 12,verify_all.pyandformal/verify_exact_rational_proofs.py: all exit 0, byte-identical, with no tracked file changed. Wave 4--checkfails on main (Canonical SHA-256 mismatch on disk) and passes here. The long Wave 7 check was not rerun.lake build(Lean 4.35.0):Build completed successfully (15 jobs).test_wave11_12_corrections.py: OK.research/update-d144448-review/implementation/check_docs.py: passes,Unrelated manifest mismatch count: 0(6 on main).Original description from #15:
Main at
d144448has 11 file hashes that disagree with its 36-entry manifest. The Wave 4 certificate also spells its canonical hash key assha_26_canonical, causing the existing generator check to fail.Restore the expected
sha256_canonicalkey without changing its value or any scientific field. Refresh all 36 hashes and record the current audit alongside the historical snapshot inCHECKSUM_PROVENANCE.md.Validation on the refreshed branch:
--checkpasses.0bf72bf.Historical validation at
0bf72bf: nine generator checks (Waves 3, 4, 5, 6, 7, 9, 10, 11, and 12) and the two checks inverify_all.pyandformal/verify_exact_rational_proofs.pypassed. Those 11 checks are earlier evidence. The unchanged long Wave 7 check was not rerun for this documentation update.This manifest records bytes. It does not validate physical models or prove that Lean files compile. The full Lean build needs the separate corrections in PRs 10 and 14. Recompute hashes when any listed file changes.