Cyan Eclipse Verbatim — Executive Orders

Vendor: Cyan Eclipse Technologies · Platforms: MCP over stdio

Certification scope

CERTIFIED(local) — Eclipse Bazaar Test Bench 2026-10-26, standing card eclipse-verbatim-eo-mcp-2026-10-26-3.json (scoped re-bench #3 of fix round 2, commit 56537e4; cards -1/-2 retained append-only — the mechanism was certified-grade on card -1 and certification was withheld on card -2 on EO-D1-2 alone). Proven: the 10-tool local stdio surface as shipped (builder suite 76/76 in substance — 72 PASS on plain bench runs plus 4 copy-dependent tests PASS under the bench replication harness, their plain-run failures being this VM's environment only; smoke PASS re-run by the bench); snapshot eo-20261022-26818c895dee (graph-3): 1,315 sealed documents, 10,667 edges, 823 divergence records, all OPEN (OMISSION 719, TYPE_DISAGREEMENT 104). EO-D1-2 CLOSED — the one defect that ever falsified a served claim: whole-cue scope emission trusted the savings idiom (a continuation assertion) and partial strikes; the graph-3 FULL census is 9 → 5, every surviving FULL re-derived by the bench's own reading of the sealed asserting text, the 4 unsupported FULLs corrected exactly (EO 11958→11501 twice and EO 13637→11958 to UNSPECIFIED; EO 13470→12333 to IN_PART 'Part 1'), and an exhaustive sentence-level sweep of all 13,282 sealed segments finds exactly 3 FULL candidates — precisely the genuine TEXT units; the idiom strip never eats a genuine cue (overreach battery: both orders, replacement form, single-clause continue+revoke). EO-D1-3 CLOSED: all 186 stated ranges re-derived verbatim in their pinned segments (186/186), binding sample 27/27. EO-D8 CLOSED in substance: graph-2 → graph-3 edge sets identical (10,613 unique keys), exactly 6 scope-only diffs, documents_root + divergences_root byte-identical, zero assertion groups gained or lost, net −66 exact (observation EO-O9: the spec A6 −94/+28 instance split differs by one each way from the bench's grouping, −93/+27 — a counting-convention artifact; a future accounting should state its grouping key). Riders CLOSED: EO-D5-2 (9/9 fresh Public Inspection marking variants refuse PUBLIC_INSPECTION_NOT_INGESTED by name against an empty PI index; bare 'PI' stays generic −32602, never stretched) and EO-D7 (§8.3 snapshot + coverage blocks on all 10 tools' successes and on refusal envelopes, SNAPSHOT_NOT_FOUND and GRAPH_VERSION_MISMATCH standing without blocks by documented construction; disposition carries divergences[] and self_limiting_terms[] as serve-time-derived, hash-exact pointers, never evaluated). Extraction bar re-measured on a fresh 15-assertion labelled sample: precision 1.000, recall 1.000; corpus FULL precision exhaustive at 5/5. NEW LOW FINDINGS (synthetic-only, zero sealed instances, corpus-swept): EO-D1-4 — four whole-cue binding residuals outside the implemented grammar (truncated idiom without 'and effect'; roman-numeral parts; an appendix struck wholly; a comma-joined cue about another object), recorded with a PRE-RESEAL CONDITION — the next reseal's bench sweeps the new batch for these shapes before the snapshot serves as current; EO-D1-5 — hyphenated compound ordinals ≥13 misquote the range tail ('twenty-first section' → range 'first section'), a verbatim-accuracy gap, never FULL. Observation EO-O8: snapshot history holds two builder-iteration seals (d31caba, ede0001) and graph-3's manifest names ede0001 as its previous rather than c44a7ad80308 directly — append-only history is honestly kept and the certified lineage reconciles exactly. THE FENCE (gate C3, load-bearing): no output of this product characterizes legal effect; every disposition answer remains an edge set + as_of + coverage window + the §2.2 completeness sentence, bench-verified byte-identical to the spec constant. Coverage re-derives from the manifest: sealed from 2000-02-10, 258 exclusions all NO_FR_FULL_TEXT_XML (257 documents 1994–1999 + EO 13548), unsealed layer from 1931-08-26; 1,315 + 258 = 1,573. SIGNING: the snapshot seal is a FIXTURE seal (FIXTURE_SIGNED) — never production Herald custody. NAME CONDITIONAL: the name inherits the pending Eclipse Verbatim trademark screen (owner-side); certification covers the local product's function under its working name, not the name.

What it does

Sealed corpus, recomputable by anyone: 1,315 Executive Orders (Federal Register full-text XML) sealed per section under canon-eo/1.0 into snapshot eo-20261022-26818c895dee — per-section hashes, a signed manifest, append-only snapshot history; the bench re-derives the seal from the stored artifacts and a spec-only build-time recompute converged 1,315/1,315 document hashes identical
The novel mechanism — a disposition graph extracted from the orders' own sealed sections: 10,667 edges, every edge pinned to the hash of the asserting section and typed by lineage (TEXT / FR_NOTE / NARA_TABLE), so a graph claim can always be recomputed against the sealed text that asserts it; where the consulted sources disagree, the disagreement itself is sealed as a DIVERGENCE record (823, status OPEN forever, never resolved by the product)
Scope is clause-bound, never idiom-bound: a whole-order cue binds only in a clause naming the target; part-governed cues mint IN_PART with the stated range; the savings idiom is stripped before cue search and can never mint a whole-order scope — the bench's exhaustive sentence-level sweep of all 13,282 sealed segments finds exactly the 5 genuine whole-order edges and zero contamination
Honest coverage, measured not assumed: the Federal Register full-text XML floor is 2000 (the 1994 gate assumption was falsified by measurement) — 258 orders are EXCLUDED_NAMED (NO_FR_FULL_TEXT_XML), never silently absent; pre-2000 orders ride in an unsealed metadata+edges layer reaching back to 1931, labelled unsealed on every surface; Public Inspection versions are refused by name (PUBLIC_INSPECTION_NOT_INGESTED)
The fence, by construction (gate C3): no tool characterizes legal effect and no status field exists anywhere in the product — every disposition answer is an edge set + as_of + a coverage window + the standing completeness sentence, served byte-identically; cite_check verifies a claimed section against its sealed hash, and resolve_authority resolves an order's statutory citations against the consumed sealed Code corpus (an empty port answers by name until that corpus lands — never a fabricated resolution)

Tools

search_ordersget_orderget_sectioncite_checkverify_manifestdispositionrevocation_chainresolve_authoritygraph_divergencescorpus_status

Problems it addresses

I need the exact text of an Executive Order section with a hash I can verify — not a scrape I have to trust
The order's own text, the FR disposition notes, and the NARA tables tell different stories — I need the disagreement preserved as sealed data, not resolved away
I need what the named sources assert about an order, as graph state as of a named snapshot — without anyone dressing graph state up as legal effect

Not tested

Hosted or remote serving — v0.1 is local stdio only; no hosted serve path exists or was certified.
Production Herald custody — the snapshot seal is FIXTURE_SIGNED with a labelled fixture key; no production key was exercised.
Sealed texts before 2000 — the Federal Register full-text XML floor is 2000 (measured); 1994–1999 orders exist only in the unsealed metadata+edges layer, and no sealed text is claimed for them.
Any characterization of legal effect — refused by construction: no status field exists; disposition answers report what the named sources assert, as graph state as of a named snapshot, with the standing completeness sentence.
EO→USC authority resolution against a live Code corpus — the port is empty by design at v0.1 and answers by name (CODE_CORPUS_UNAVAILABLE); resolution activates when the sealed Code corpus lands.
The trademark screen for the Verbatim family name (owner-side) — still open; it gates public surfaces under the name, not this local certification.
Any reseal after card -3 — EO-D1-4 / EO-D1-5 are recorded with a pre-reseal sweep condition; no reseal has occurred since certification.

Bench scorecard: test-bench/results/eclipse-verbatim-eo-mcp-2026-10-26-3.json

Evaluation-intake listing

This listing records the tool's place in the Bazaar's evaluation and monitoring pipeline, with the evidence gathered so far printed above — verdict, scope, and not-tested items included. It is not sold on this site, and its presence in the catalog is not a claim beyond that evidence. The service behind every verdict is the bench: what a bench run buys →