Repository navigation
Remeasure retained 1SAR models with current evidence contracts - #1007
realmarcin wants to merge 3 commits into
Conversation
Bind the four archived subjects and shared observations to exact identities, define a two-producer R-factor canary, and require retained selections and current EVAL/QDS evidence before expanding the run. State directional reproduction predictions, inference limits and unanswerable outcomes. Record the pre-execution corrections for input aliases, API/catalog references and distinct clashscore occupancy policies (#1004, #1005, #1006). This commit contains no new scientific results or tool execution.
|
Additional in-scope input finding: #1018. An independent byte-level check found20 excluded rows with nonfinite F/SIGF and missing flags in the exact retained MTZ. The draft ledger rejects flags before observation eligibility and therefore rejects the preregistered input. The fix must preserve these raw rows and distinguish unassigned excluded flags, while keeping usable-observation invalid flags fatal. No scientific measurement was run; this is queued with the evidence corrections before the canary. |
|
Fresh read-only reviews102/103 and PRIMARY verification found further in-scope work before the canary:
PRIMARY reproduced the two admission gaps using synthetic fixtures; no scientific measurement ran. Focused suite:17tests,1failure/1error; lint fails at the named expression. All findings are in scope because they affect the mandatory gate or registered evidence contract. Fixes will be dispatched in smaller tasks, then independently verified and re-reviewed. No merge requested or authorized. |
|
Read-only implementation planning identified one additional documentation finding: #1022, the current evaluation skill's incorrect claim that4σ is the installed PHENIX default. Triaged in scope as a bounded prose correction alongside the optional peak-inventory methods. The preregistered±4σ setting and all historical observations/assumption ancestry remain unchanged. No scientific run has occurred; the PHENIX canary collector remains under local correction and verification. |
|
Two further source-verified documentation findings are linked here: #1023 (T05 driver overstates Reduce/Probe classifier capabilities) and #1024 (it labels deposited1SAR1.20Å instead of the retained1.80Å). Both are triaged in scope as small prose corrections; the scientific protocol, thresholds, catalog identities and historical records remain unchanged. The current canary implementation is still under local verification, with no scientific run yet. |
Stage the exact archived inputs and record raw, eligible, extracted and final reflection populations. Bind the configured measured build and vendor sources, replay native R factors from retained arrays, and verify scatterer invariance, runtime symmetry and serialized map supplements. Retain failed-stage logs and raw evidence without admitting failed or unmeasured results. Add hermetic source-backed and cache-boundary tests. This phase contains no new scientific measurements; the independent collector and record-authoring CLI follow separately.
|
Fresh independent review of the preparatory REFMAC parser/transport found the following issues. PRIMARY reproduced the defects and independently verified the retained mutation inputs, outputs and hashes. All are in scope before the canary; none is a new scientific result.
The percentage fix is limited to its documented [0,100] domain: exact same-cycle denominator arithmetic remains unproven from installed sources. Exact Decimal equality may preserve different printed trailing precision; no approximate numeric tolerance is being introduced. The accepted PHENIX phase remains at af5fb0c. The current parser candidate has not passed QC or been committed, and the scientific canary has not run. |
|
The independent transport-only focused check found a newly introduced syntax failure before any test executes: #1038 ( |
|
Fresh independent production re-review found three bounded parser defects, reproduced by PRIMARY against a pinned historical native log. The new lexical-token finding is #1039; graph-cycle contradictions are tracked under existing #1032 (related #1034), and exponent-precision metadata under #1036. All are in scope. Remaining parser fixture failures are recorded under #1026. No scientific tool run or canary has occurred. The corrected transport suite passes, and a separate fresh transport review found no actionable defects. |
Add a pure parser that preserves native tokens, counts, precision and source lines while reconciling ordinary R across detailed, summary and final reports. Reject contradictory cycle identities, malformed numeric values and incoherent reported counts. Extend the shell-free wrapper with explicit executable, environment and XML-output handling and retain legacy defaults. Hermetic parser and fake-child transport fixtures cover the rejection paths. This prepares the 1SAR collector; it contains no new scientific measurements. Refs #1026, #1027, #1028, #1029, #1030, #1031, #1032, #1033, #1034, #1035, #1036, #1037, #1038, #1039.
|
Accepted and pushed parser/transport preparation in b8017ba. The actual20-test parser suite, transport fake-child suite,24 retained diagnostic cases, required locked hermetic gate and Ruff pass. Separate fresh transport, production-fix and final-test reviews have no actionable findings. This addresses #1026–#1039, including the deduplicated cycle/precision/fixture follow-ups. Native historical parsed output remains byte-identical after the fixes. This is infrastructure only: no new 1SAR scientific tool run or canary has occurred. Parsing does not establish effective HKLs, residual reconstruction or model preservation. Next is the collector and one-subject CLI/records path. New-head CI is queued; the PR remains draft and unmerged. |
|
Fresh read-only collector reviews170/171 and PRIMARY synthetic diagnostics found issues in the uncommitted output-audit candidate (SHA-25645200f3f20258cea5f6506cf9fed5eca4e5e991b08ca1f92f8bfcf106005f4be on parentb8017ba). New findings: #1040, #1041, #1042, #1043, #1044, #1045, #1046. Deduplicated recurrences: #1008, #1030, #1032. All are in scope before the single R4 canary. PRIMARY verified helper API failures, ledger/partition errors, XML numeric/precision/association behavior, failed-attempt persistence, temporary-file ownership and cell-format policy against actual code and retained synthetic artifacts. The candidate is not accepted. No vendor tool, refinement, benchmark or new scientific measurement has run. The current pushed head and its green checks remain b8017ba; corrections, tests, full QC and fresh re-review precede any new commit or canary. Historical records remain unchanged. |
|
Follow-up acceptance of the reported collector patch confirms the work/free partition arithmetic and PDB helper adapter, but the complete MTZ audit still fails. Remaining verified fixes are recorded on #1008 (binding return shape), #1040 (required count validation) and #1046 (exact CELL/DCELL numerical equality). All remain in scope before the R4 canary; the collector is not accepted or committed yet. No scientific run occurred. |
|
The bounded Task175 retention correction passes the retained audit-failure, missing-input and collision checks, including exact final audit artifacts. A fresh read-only review and PRIMARY synthetic reproductions found two remaining P2 checkpoint cases: undetected modification/removal of the pre-execution checkpoint, and unreadable output inspection escaping before post-execution/final records. Both are deduplicated under #1044 and remain in scope before the R4 canary. See #1044 (comment). A narrow follow-up fix is prepared. No scientific execution or new measurement has occurred; the collector is not yet accepted. |
|
The Task192 leaf-checkpoint and unreadable-file corrections are independently verified, with prior audit-retention checks still passing. A fresh-context review and PRIMARY reproductions found two remaining directory-level cases: failed final persistence when the child directory is non-writable despite a writable evidence root (P1), and false original-path checkpoint verification through an ancestor symlink (P2). Both are deduplicated under #1044, remain in scope before the canary, and are detailed at #1044 (comment). The next bounded fix adds safe parent retention and complete ancestor/directory verification. No scientific execution or new measurement has occurred; full collector acceptance remains pending. |
|
Fresh-context retention review204 and PRIMARY reproductions identified five remaining in-scope cases, deduplicated on #1044: #1044 (comment). These cover rejected input/root traversal, parent-retention identity, late persistence status and the saved manifest self-path. Earlier bounded retention cases pass, but the collector remains unaccepted until these fixes and re-review pass. No scientific canary or vendor execution has occurred. |
|
Pre-commit review of the current local REFMAC final-pair candidate found five in-scope defects. The new collision-state finding is #1049; the nonlocal-scope recurrence is recorded on #1038, and three remaining terminal/publication-state cases are recorded on #1044. All must be corrected before the R4 canary. PRIMARY confirmed that frozen local source63fc67d24bd897927258165eeaa9f58d49b53b91eb58fecf6117c522c0b1dfcf does not compile under Python3.12.11; the other findings are source-confirmed downstream paths, not claimed runtime results. This collector remains uncommitted and is not covered by CI on the current pushed headb8017ba0af89edfb1d8081bb71380708ff9554d0. No vendor run, benchmark or scientific canary has occurred. The secondary will first make the bounded state-name repair; PRIMARY owns execution, verification and QC. No merge is authorized. |
|
Fresh review273 of the latest local collector correction found no introduced regression. It identified the unchanged concurrent directory-replacement race recorded in #1050. Triage: follow-up, kept open; it requires an external replacement after admission and a separate descriptor-bound publication contract. The reviewed sequence ends incomplete and does not grant scientific admission. The in-scope sequential terminal-state fixes on #1044 passed PRIMARY verification of six synthetic controls and 55 actual saved mappings. The source remains uncommitted while Task176 numeric fixes and permanent tests/QC proceed. Existing green checks apply to pushed head b8017ba, not these local edits. No merge or scientific measurement is claimed. #1050 is not an intended closure of this PR. |
|
The fresh numeric-policy review of the uncommitted REFMAC collector is not yet accepted. PRIMARY verified the existing27 controls plus five focused reproductions, with all vendor tools disabled. Remaining in-scope findings are recorded on 1042 (exact interval arithmetic and unsupported source-level contradiction), 1037 (original-token identity), and 1046 (incorrect source/selection provenance). The secondary is making the bounded corrections in wt-078; PRIMARY will verify them, re-review, run the permanent tests and QC before committing. Current pushed head remains b8017ba. No scientific canary has run, and existing green CI covers only that pushed head. The separately triaged concurrency follow-up1050 remains open. |
This draft prepares a reproducible remeasurement of the retained 1SAR trajectory. The published no-results preregistration pins the actual start, packaged R4, retained R5 and R6 inputs by SHA-256. R7 remains log-only because its coordinates were not retained.
The implemented PHENIX collector stages those exact inputs, binds the configured measured build and vendor sources, retains separate input/extracted/effective reflection populations and failed-stage logs, replays native R factors from full-precision arrays, and verifies scatterer invariance, runtime symmetry and the written map supplements. Invalid or unmeasured evidence cannot become an accepted result.
This is still a preparation-only draft. The independent REFMAC collector, end-to-end canary/EVAL/QDS path, four-subject measurements and dated assessment remain pending. No scientific tool, refinement, benchmark or data download has run for this study. Historical records are unchanged.
Validation of this adapter phase: PRIMARY's actual focused suite (19 tests), Ruff and the full locked Python 3.12 benchmark-extra hermetic gate; fresh read-only reviews of the implementation fixes and test corrections; independent synthetic checks of retained failure artifacts and symmetry rejection. These checks validate the adapter implementation, not a scientific result or the unimplemented independent producer.
Review tracking: #1004 and #1005 concern the registered inputs/APIs; #1006 governs the later clashscore adapter. The PHENIX implementation findings are #1008, #1009, #1010, #1011, #1012, #1013, #1014, #1015, #1016, #1017, #1018, #1019, #1020, #1021 and #1025. Related remaining work includes the dated assessment correction #1003 and bounded documentation findings #1022, #1023 and #1024. Issues remain open while this draft analysis is completed.
The accepted analysis records will supply the new slide triplets in deck PR #998. Neither PR is authorized to merge. This description and the validation record will be updated for the completed implementation and actual evidence before final review.