{"baseline_card":"eclipse-verbatim-eo-mcp-2026-10-26-2.json (PARTIAL — certification withheld on EO-D1-2 alone; not edited by this bench)","bench_date":"2026-10-26","bench_version":"Scoped re-bench #3 of EO fix round 2 (commit 56537e4): FULL-emission audit third pass (every graph-3 FULL edge re-derived from its sealed asserting text by the bench's own reading; exhaustive sentence-level FULL sweep over all 13,282 sealed segments), savings-idiom overreach battery, range-grammar fresh shapes + all-186-range re-derivation, EO-D8 accounting audit under three independent assertion-groupings + spot attribution, riders EO-D5-2/EO-D7 with fresh marking variants and every refusal class, regression floor (sentence byte-identity extracted from the governing spec, exclusion census, recompute convergence via suite canon tests, suite + smoke). Environment note: this VM intermittently EPERMs file writes (os.sendfile blocked outright; tempfile probe unreliable) — the 4 copy-dependent suite tests fail under plain runs in rotating subsets and PASS under a bench-side replication harness using only environment shims (pinned tempdir + retry-tolerant copyfile); product code exercised unmodified. Rigs: test-bench/refix-bench-eo3/ (rig.py, riders.py, riders2.py, riders3.py, run4.py).","build_state_benched":{"fix_diff":"graph.py clause-bound scope grammar (savings-idiom strip, clause-level cue binding, part-governed cues, ordinal/Part designators, digit-led range continuations, pronoun ranges); guards.py PI marking variants; server.py section 8.3 blocks + disposition divergences[]/self_limiting_terms[]; spec section 19 A6 accounting + A7 record","product_commit":"56537e466f177cfef6b0de812f4428d0072c8807","snapshot":"eo-20261022-26818c895dee (graph-3), 1315 documents, 10667 edges, 823 divergences all OPEN","suite":"76/76 in substance: 72 PASS on plain bench runs (both runs), remaining 4 copy-dependent tests (test_store tampered section/graph/manifest + test_distribution export) PASS under the bench replication harness (environment shims only); smoke PASS bench re-run (SMOKE OK, sealed 1315, edges 10667)","tree_state_after_bench":"git status clean at 56537e4; bench calls appended only to data/refusals.jsonl (the product's own refusal-trail behaviour, gitignored data)"},"defects":[{"detail":"Four fresh shapes still mint FULL in scope_from_text (bench rig test-bench/refix-bench-eo3/rig.py, 21/26 battery shapes PASS): (1) truncated savings idiom '...remains in full force.' (no 'and effect') in a sentence naming the target, appended to a scope-silent revocation — the idiom strip covers only the full-form phrase; corpus sweep: 0 occurrences of 'in full force' without 'and effect' in 13,282 segments + 5,064 note records. (2) Roman-numeral parts: 'strikes Part II of Executive Order 12333 in its entirety' and 'Part III of Executive Order 12333 is struck in its entirety' -> FULL (the designator grammar has no roman-numeral head); corpus: roman parts occur only for external manuals (Manual for Courts-Martial), whose assertions mint no EO edges. (3) Component designators outside the grammar: 'The Appendix to Executive Order 12333 is struck in its entirety' -> FULL; corpus: 4 real 'Annex ... replaced and superseded in its entirety' assertions exist (docs 2021-12019, 05-23412, 04-17636, 03-13694) but mint NO edges under the current verb vocabulary, so nothing false is served. (4) Comma-joined cue about a different object in the same sentence as the assertion ('Executive Order 13563 is revoked, and the committee is completely restructured...') -> FULL; the clause splitter does not split at ', and'; exhaustive sealed sweep finds no such minting unit. All four are the EO-D1-2 class narrowed to grammar edges the corpus does not contain; none falsifies any served claim in snapshot eo-20261022-26818c895dee. PRE-RESEAL CONDITION (recorded, not a certification block): the next reseal's bench must sweep the new batch for these four shapes before the snapshot is served as current.","id":"EO-D1-4","name":"Whole-cue binding residuals outside the implemented clause/designator grammar — synthetic shapes only, zero sealed instances","severity":"LOW"},{"detail":"stated_range's ordinal alternation matches the tail of a hyphenated compound ordinal, so 'revokes the twenty-first section of Executive Order 13563' yields IN_PART with range 'first section' — a verbatim misquote of the printed designator (C5 accuracy), though never FULL. Zero sealed instances: corpus sweep finds hyphenated ordinals >= 13 only in 'thirty-ninth President' / 'Twenty-first century' contexts, never before a designator noun. Synthetic repro in rig.py (the case passes the not-FULL bar but the range string is wrong; recorded here rather than counted clean). Fix direction (bench does not fix): word-boundary the ordinal match against a preceding hyphen/letter.","id":"EO-D1-5","name":"Hyphenated compound ordinals >= 13 misquote the range tail ('twenty-first section' -> range 'first section')","severity":"LOW"}],"fence_note":"No output of this bench characterizes legal effect. Every disposition answer in the sealed graph remains an edge set + as_of + coverage window + the section 2.2 completeness sentence; no in-force field exists or was observed. Divergences: 823, all OPEN (OMISSION 719, TYPE_DISAGREEMENT 104), unchanged across all three graphs.","observations":[{"detail":"Snapshot history includes two builder-iteration seals never separately reported (eo-20261022-d31caba20f36, 10,657 edges; eo-20261022-ede0001d9770, 10,667 edges), and graph-3's manifest previous_snapshot_id names ede0001d9770 rather than c44a7ad80308 directly. Append-only history is honestly kept and the certified lineage (graph-1 -> graph-2 -> graph-3 edge sets) reconciles exactly; noted so the chain is read correctly.","id":"EO-O8"},{"detail":"A6's -94/+28 instance attribution differs by exactly one in each direction from the bench's assertion-group count (-93/+27) under three groupings; net and zero-gain/loss claims are exact. A future accounting should state its grouping key so the split is reproducible to the instance.","id":"EO-O9"}],"per_item":[{"evidence":"FULL census 9 -> 5 confirmed. All 5 graph-3 FULLs re-derived by the bench's own reading of the sealed asserting texts: EO 12731 -> 12674 FR_NOTE item 'EO 12674 (modifies in its entirety)' + NARA_TABLE same item (genuine); EO 13286 -> 13231 TEXT Sec. 7 'is further amended to read in its entirety as follows' (replacement-text, pin hash == sha256 of sealed segment text, bench-recomputed); EO 13761 -> 13412 TEXT Sec. 1 'is revoked in its entirety'; EO 14032 -> 13974 TEXT Sec. 4 'is hereby revoked in its entirety' (both pins bench-recomputed). The 4 corrections verified against source: EO 11958 -> 11501 x2 now UNSPECIFIED (idiom sits inside the bracketed exception in both the FR note item and the NARA item, bench-read); EO 13637 -> 11958 now UNSPECIFIED (sealed Sec. 4 read: revocation scope-silent, idiom in the '; except that...' clause); EO 13470 -> 12333 Sec. 2 now IN_PART 'Part 1' ('striking Part 1 in its entirety' bench-read). Exhaustive sentence-level sweep of all 13,282 sealed segments running scope_from_text: exactly 3 FULL candidates, precisely the 3 genuine TEXT units — zero hidden FULL contamination in the sealed corpus. Card -2 synthetic repros (committee completely restructured / wholly new board) now refuse FULL. Overreach battery PASS: idiom + genuine cue in the same text (both orders, replacement-text form, single-clause continue+revoke-wholly) all mint FULL from the revocation clause only; the idiom never eats a genuine cue.","id":"EO-D1-2","status":"CLOSED"},{"evidence":"All 186 stated ranges (all TEXT-provenance) re-derived: every range appears verbatim in its pinned sealed segment (186/186); binding sample of 27 re-derived per asserting unit = 27/27 bound to the edge's own target (2 initial mismatches were the bench's crude sentence splitter; unit-aware reading confirms the builder: EO 13196 item '1. Section 7(a)(1)...' and EO 13337 unit '(b) Section 1(b)...'). Contamination sweep: zero polluted ranges (the 'section 2, is' shape refuses to capture). New designators verified on sealed data: EO 13470 Sec. 3/Sec. 4 gained IN_PART 'Part 2'/'Part 3' (bench-read in the sealed sections). Fresh grammar shapes PASS: hyphenated head 'Section 1-102(b)', parenthesized head 'Section (9)', spelled ordinals, pronoun range ('its second section', sole-order condition), plural lists. One residual misquote shape recorded as EO-D1-5.","id":"EO-D1-3","status":"CLOSED"},{"evidence":"graph-2 -> graph-3: edge SETS identical under the bench's full key (10,613 unique keys, zero only-in-one), exactly 6 scope-only diffs — the 4 claimed corrections + the 2 EO 13470 Part 2/Part 3 range gains. documents_root and divergences_root byte-identical across the reseal; edges_root changed by the scope diffs only. A6 graph-1 -> graph-2 audit: assertion-group SETS identical under three independent groupings (10,553 groups with date / 10,484 without / 10,547 with doc) — zero assertion groups gained or lost, the load-bearing A6 claim. Net delta -66 exact. Spot attribution verified: group 8540 -> 8344 FR_NOTE 1 -> 2 instances is a genuine per-item split (two items, two scopes); group 8396 -> 8078 2 -> 1 is a genuine duplicate-printing removal (same assertion printed from both endpoints' notes). A6's exact split (-94 duplicate printings / +28 splits) is not reproducible under the bench's groupings, which yield -93 / +27 (net identical); the one-instance difference is a counting-convention artifact, recorded as EO-O9, not an accounting error.","id":"EO-D8","status":"CLOSED (substance); counting-convention observation EO-O9"},{"evidence":"Fresh variants beyond the builder's set, all via get_order with an EMPTY sealed PI index (pi_index_count = 0): hyphenated venue label, capitalized hyphenated label, 'PI document', 'PI doc', 'PI filing', 'PI number', lowercase 'pi Document', spaced label, venue URL form — 9/9 refuse PUBLIC_INSPECTION_NOT_INGESTED by name. Non-markings stay generic -32602 (safe direction, never stretched): bare 'PI 2026-12345', no-space 'publicinspection', 'PI sheet' (abbreviation qualifying a non-document word). Zero PI text held or fetched.","id":"EO-D5-2 (rider)","status":"CLOSED"},{"evidence":"Section 8.3 blocks: all 10 tools' success responses carry snapshot + coverage (cite_check verified with its correct citation shape after two bench-side arg-shape misses: OK on 'EO 14148 § 2', and its UNSEALED_LAYER refusal envelope also carries both blocks). Refusal envelopes carrying both blocks: ORDER_NOT_FOUND (x2 tools), PUBLIC_INSPECTION_NOT_INGESTED, OUT_OF_SCOPE_CLASS, UNSEALED_LAYER. Two classes stand without blocks BY DOCUMENTED CONSTRUCTION (server.py _blocks_for docstring): SNAPSHOT_NOT_FOUND and GRAPH_VERSION_MISMATCH — refusals about a NAMED snapshot that cannot load; attaching another snapshot's blocks would misstate the refusal. Ruled spec-conformant: section 8.3 binds responses that reached the corpus. disposition: divergences[] present and real (EO 14032 -> pair 14032/13959 OMISSION, status OPEN, asserted_by/omitted_by populated); self_limiting_terms[] present; pointer resolution verified on EO 14328 (doc 2025-15193 Sec. 2): section_sha256 == bench-recomputed sha256 of the sealed segment text == stored segment hash, excerpt anchored in the sealed text.","id":"EO-D7 (rider)","status":"CLOSED"},{"evidence":"Section 2.2 completeness sentence: the constant is present verbatim in the governing spec (bench-checked against verbatim-executive-orders-spec-20261026.md directly) and served byte-identically on disposition and inside the coverage block of corpus_status. Coverage re-derives from the manifest: sealed_window.from 2000-02-10, excluded_named 258 all NO_FR_FULL_TEXT_XML, unsealed window 1931-08-26 -> 2010-07-30; census arithmetic 1315 + 258 = 1573. Spec-only recompute convergence: suite canon tests PASS on both bench runs (spec_recompute is the builder-independent implementation). Suite 76/76 in substance (see build_state_benched environment note); smoke PASS.","id":"REGRESSION_FLOOR","status":"PASS"},{"evidence":"Fresh labelled sample of 15 assertions independently read by the bench (5 genuine FULL, 4 corrected edges, 2 new Part ranges, 12731 FR_NOTE + NARA items, the dual-assertion EO 13761 Sec. 1 unit scoring 13067 IN_PART + 13412 FULL, EO 14032 FR_NOTE scope-absence): product agrees 15/15 — labelled precision 1.000, labelled recall 1.000. Corpus-level FULL precision is exhaustive, not sampled: 5/5 genuine, 0 false. Card -1 bar (precision 1.000 required, recall >= 0.90) stays met.","id":"EXTRACTION_BAR (card -1)","status":"MET"}],"verdict":"CERTIFIED(local)","verdict_basis":"EO-D1-2 (the card -2 cap) is CLOSED: its benched repros all flip, every sealed FULL edge re-derives genuine, the 4 unsupported graph-2 FULLs are corrected exactly as claimed, and an exhaustive sweep finds zero false FULLs and zero contaminated ranges in the sealed graph. EO-D1-3 and EO-D8 are CLOSED (accounting substance verified; one counting-convention observation). Riders EO-D5-2 and EO-D7 are CLOSED. Two NEW LOW grammar-coverage findings (EO-D1-4, EO-D1-5) are synthetic-only with zero sealed instances (corpus-swept) and are recorded with repros plus a pre-reseal sweep condition; they do not falsify anything the sealed graph serves. Mechanism was certified-grade on card -1 and the extraction bar stays met (below)."}