diff --git a/docs/release-1.0-criteria.md b/docs/release-1.0-criteria.md
index a64075a..99a7f7c 100644
--- a/docs/release-1.0-criteria.md
+++ b/docs/release-1.0-criteria.md
@@ -1,6 +1,6 @@
 # 1.0.0 release criteria
 
-Last measured 2026-10-21, on `main` at `b268e884`.
+Last measured 2026-10-27, on `main` at `61255e9` ([CI run 40542129639](https://github.com/DrunkOnJava/dwg-rs/actions/runs/40542129639), artifact `corpus-measurement`).
 
 1.0.0 is the first release with a stability commitment
 ([`RELEASE.md`](../RELEASE.md)). The maintainer set its scope on
@@ -27,13 +27,13 @@ status changes only in the pull request that links the new evidence.
 | | Requirement | Bar | Measured now | Status |
 |---|---|---|---|---|
 | R1 | Measured in public | Every number README.md, STATUS.md and this page publish comes from a CI run on files anyone can fetch, with the run linked | README.md's and STATUS.md's real-file numbers come from the Corpus CI job on 52 public drawings, run linked; STATUS.md's other numbers and this page's do not yet (#94, #112) | **failed** |
-| R2 | Every release opens | R13, R14, R2000, R2004, R2007, R2010, R2013 and R2018 files open and walk, on at least two corpus files per release | R14 to R2018 open and walk in CI on at least four drawings per release (eight each for R14 to R2013, four for R2018, all AutoCAD saves with recorded provenance); no R13 file (#119) | **failed** |
-| R3 | Every section of the spec is read | Every row of [`spec-coverage.md`](spec-coverage.md) is `complete`: chapters 2 to 29 and all 104 prescriptions of §20.4 | 0 of 138 rows `complete`; 34 `missing`, 31 `partial`, 6 `decoder`, 67 `closes` (CI corpus, #225) (#118, #125, #131, #138) | **failed** |
-| R4 | Values equal AutoCAD's | For every section, every value equals AutoCAD's DXF export of the same drawing, matched by handle | The DXF value oracle (`examples/dxf_oracle`, #99) compares the decoded fields of 46 sections on R2000-R2018 (entities also on R14) over the nextgis and ACadSharp exports in the Corpus CI job, which fails when an agreeing count falls below `tools/oracle/baseline.tsv` (#103): 19 values disagree (#241-#244), every other decoded value agrees; most sections still have DXF groups no decoder surfaces, and no R13 drawing is in the corpus | **failed** |
+| R2 | Every release opens | R13, R14, R2000, R2004, R2007, R2010, R2013 and R2018 files open and walk, on at least two corpus files per release | R14 to R2018 open and walk in CI on at least four drawings per release (eight each for R14 to R2013, four for R2018, all AutoCAD saves with recorded provenance; 52 drawings, 15,541 records decoded, 0 errored, `corpus-results.json` of [the run](https://github.com/DrunkOnJava/dwg-rs/actions/runs/40542129639)); no R13 file (#119) | **failed** |
+| R3 | Every section of the spec is read | Every row of [`spec-coverage.md`](spec-coverage.md) is `complete`: chapters 2 to 29 and all 104 prescriptions of §20.4 | 0 of 138 rows `complete`; 28 `missing`, 7 `partial`, 6 `decoder`, 97 `closes` (`docs/data/spec-coverage.tsv` with #398, whose seven moves to `closes` and §22 to `partial` are proven by [CI run 40532346172](https://github.com/DrunkOnJava/dwg-rs/actions/runs/40532346172); `closes` is measured by the Corpus CI job, #225) (#118, #125, #131, #138) | **failed** |
+| R4 | Values equal AutoCAD's | For every section, every value equals AutoCAD's DXF export of the same drawing, matched by handle | The DXF value oracle (`examples/dxf_oracle`, #99) compares the decoded fields of 53 sections on R2000-R2018 (entities also on R14) over the nextgis and ACadSharp exports in the Corpus CI job, which fails when an agreeing count falls below `tools/oracle/baseline.tsv` (#103): 22 values disagree (#241-#244, #381; `oracle-summary.md` of [the run](https://github.com/DrunkOnJava/dwg-rs/actions/runs/40542129639)), every other decoded value agrees; most sections still have DXF groups no decoder surfaces, and no R13 drawing is in the corpus | **failed** |
 | R5 | Classes beyond the spec | Every class the corpus holds that v5.4.1 does not prescribe (VISUALSTYLE, MATERIAL, surfaces and the rest) meets R3 and R4 | VISUALSTYLE and the view styles close; the DXF oracle finds every VISUALSTYLE value of 1,008 records (R2000-R2018) equal to the export but `format_version` on R2013-R2018 and Realistic's face lighting quality on R2000-R2007 (264 values, #250), and does not score the view styles yet; MATERIAL, PROPERTYSET_DATA, MESH and others partial or unmeasured (#152) | **failed** |
 | R6 | A corpus that holds everything | Every (section, release) pair has at least one corpus record with a DXF export, from sources with a documented right to fetch them | Two public sources fetched in CI: nextgis (7 simple drawings, R14 to R2013) and ACadSharp (one dense drawing, R14 to R2018, plus three R2018 drawings); AutoCAD's DXF beside every save but R14 and two of the extra R2018 drawings; no R13 (#153) | **failed** |
 | R7 | Held-out files | At least two files per release band, used by no development work and measured once before any fix, meet R3 and R4 | None reserved (#158, #196) | **not measured** |
-| R8 | Nothing silently wrong | A record that does not decode is reported (diagnostics name its type and why); no decoder returns plausible values for a layout it did not verify | 308 records error honestly on the public corpus (#226, #227, #228, #71); the DXF oracle found five decoders that closed on their boundaries with wrong values, all fixed (#235, #236, #238, #251, #253); AppInfo is misread (#137) | **failed** |
+| R8 | Nothing silently wrong | A record that does not decode is reported (diagnostics name its type and why); no decoder returns plausible values for a layout it did not verify | 0 records error on the public corpus and 15,541 decode (the 308 that errored at #237 were fixed: #226, #227, #228, #71), and 533 are skipped for want of a decoder, each named by kind in the skip histogram (`corpus-results.json` of [the run](https://github.com/DrunkOnJava/dwg-rs/actions/runs/40542129639)); the DXF oracle found five decoders that closed on their boundaries with wrong values, all fixed (#235, #236, #238, #251, #253); AppInfo is misread (#137) | **failed** |
 
 ## Writing
 
@@ -47,7 +47,7 @@ status changes only in the pull request that links the new evidence.
 
 | | Requirement | Bar | Measured now | Status |
 |---|---|---|---|---|
-| X1 | DXF | DWG to DXF equals AutoCAD's own DXF of the same drawing, for every target version the source supports | Written for R12 to R2018; checked only syntactically (#169) | **not measured** |
+| X1 | DXF | DWG to DXF equals AutoCAD's own DXF of the same drawing, for every target version the source supports | Written for R12 to R2018. The HEADER, TABLES, ENTITIES and BLOCKS sections each have a probe in the Corpus CI job that pairs every written record with AutoCAD's export of the same drawing by handle and compares it group by group; on the 42 drawings with an export of their release (10 are skipped for want of one), 0 groups disagree and none is written that the export omits: HEADER 9,583 variables (9,410 agree, 173 hold a stated relation), TABLES 1,239 records (8,831 of 8,840 groups agree, 9 by a stated relation), ENTITIES 803 records (5,627 of 5,627 groups), BLOCKS 1,532 records (12,577 of 12,583 groups, 6 by a stated relation, #243; `dxf-*-probe.log` of [the run](https://github.com/DrunkOnJava/dwg-rs/actions/runs/40542129639)). Not equal yet: 342 ENTITIES records of 18 kinds are not written (DIMENSION 66, MULTILEADER 90, HATCH 48, TOLERANCE 18, MLINE 18 and others) and the OBJECTS section is not written on `main` (#169) | **failed** |
 | X2 | SVG and PDF | Rendered geometry equals the decoded geometry for every 2D entity type, within a stated tolerance | Rendered; not measured (#171) | **not measured** |
 | X3 | glTF | Every export passes the Khronos validator; solids drawn from their ACIS data | 3DSOLID is a box placeholder (#172, #173) | **failed** |
 | X4 | Viewer | No stubs; browser tests open every corpus file; zero network requests after load | Selection, section box, clipping, worker and progressive open are stubs (#174) | **failed** |
@@ -59,7 +59,7 @@ status changes only in the pull request that links the new evidence.
 |---|---|---|---|---|
 | S1 | The stable surface is named | RELEASE.md lists what 1.0 commits to and what it does not | Every module public; nothing named (#192) | **failed** |
 | S2 | Breaking changes are caught | cargo-semver-checks and the Python API listing fail CI on an unannounced break | Not checked (#192) | **failed** |
-| S3 | Hostile input | One ParseLimits holds every cap; every fuzz target has run an hour with no findings; no panic on 64- or 32-bit targets | Separate limit structs; no recorded fuzz durations; no 32-bit job (#184) | **failed** |
+| S3 | Hostile input | One ParseLimits holds every cap; every fuzz target has run an hour with no findings; no panic on 64- or 32-bit targets | Separate limit structs; no recorded fuzz durations; the library compiles for `wasm32-unknown-unknown` in the Clippy job (step "The library builds where usize is 32 bits", success in the run), but no job runs tests or a corpus walk on a 32-bit target (#184) | **failed** |
 | S4 | Published | crates.io, docs.rs, PyPI and the release binaries checked from what was published | Nothing published (#109, #182) | **failed** |
 
 ## Outside evidence