CyanEclipseTechnologies/eclipse-test-bench-scorecards

Eclipse Test Bench Scorecards

The complete, append-only evidence corpus of the Eclipse Test Bench (Cyan Eclipse Technologies LLC): every bench scorecard issued for the MCP servers, connectors, and tools in the Eclipse Bazaar catalog — 227 files covering 98 products, including every FAILED and PARTIAL card ever issued. Nothing is cherry-picked, edited, or retired: a product's failures are part of its public record, and this corpus is that record in machine-readable form.

The live catalog these cards grade is at eclipsebazaar.com; the grading methodology is summarized below and is being published in full as the K-Check methodology specification.

Contents

File	What it is
cards/	All 227 bench files, byte-identical to the issued originals: 217 scorecards, 9 bench-evidence files, 1 dogfood note.
scorecards.jsonl	The viewer table: one row per bench file with a single fixed schema (product, bench date, verdict, scope, score, catalog cross-reference, and the SHA-256 of the card file itself). This is the config the dataset viewer renders.
index.jsonl	One derived row per card: product, bench date, verdict, scope, score (where the card records one), and the catalog cross-reference. Fields the card does not carry are null — never invented.

The raw files under cards/ are the evidence of record and stay byte-identical. They are deliberately not a viewer config: card schemas evolved across bench versions (142 distinct top-level fields across the corpus, with shared fields changing type between bench generations), so no single table schema can cast them all — the normalized scorecards.jsonl table exists precisely so the viewer renders every row while the originals remain untouched. | catalog-cross-ref.json | The Eclipse Bazaar catalog side: 83 products, the 79 carrying a scorecard reference, and whether that reference resolves to a card in this corpus (79/79 do). |

Verdict vocabulary (closed)

Verdicts use a fixed vocabulary, and the scope qualifier is part of the verdict — it may never be detached:

Verdict	Meaning
CERTIFIED(scope)	Every gating check passed at the stated scope.
PARTIAL	Some scopes or check classes passed; at least one gating item failed or remains unexercised. The failed items are named on the card.
FAILED	A gating defect was demonstrated. The product carries FAILED standing until a fix round and re-bench produce a new card.
NOT CERTIFIED	Certification was withheld (ceiling scope unreachable, or evidence insufficient to grade).

Scope qualifiers name the strongest environment actually exercised — local, simulator, staging, live, live-read, sandbox, fixture, and split-scope forms where a product's surfaces genuinely differ. A claim may only ever be made at or below the certified scope: fixture evidence never supports a live claim, and simulator evidence never supports a real-world claim.

As issued, the 217 verdict-bearing cards in this corpus read: 88 CERTIFIED at their stated scopes, 35 PARTIAL, 92 FAILED, 1 NOT CERTIFIED, and 1 split-scope CERTIFIED-at-scopes. A corpus in which most first benches fail is the methodology working as intended.

How grading works (methodology summary)

• Independent bench. Cards are issued by the certification desk (Verity), which audits and never fixes; builders never grade their own work.
• Check classes include the core-mechanism battery, adversarial fresh-shape attacks, byte-for-byte regression replication of prior defects, attacks aimed at the fix code itself after a fix round, independent recompute from the published specification alone, envelope hygiene, closed refusal vocabularies, claim discipline, and dogfood use inside the certified scope.
• The N-Check (novelty gate). Every product must state its novel mechanism. If ≥80% of the shipped surface is renamed incumbent vendor verbs, the N-Check fails regardless of prose: the verdict is capped and the novelty claim is struck from all copy.
• Fix rounds. A FAILED or PARTIAL product may be fixed and re-benched; regressions must replicate the bench's shapes byte-for-byte, and each round issues a new card. No round erases a prior card.
• Claim discipline. Scope words travel with the verdict on every surface — catalog cards, registry descriptions, product pages. "CERTIFIED" without its qualifier is a claim defect.

Recompute rights

Anyone may re-derive any claim in this corpus: the cards carry their own check evidence, the products' specifications are published with the products, and the catalog verdicts at eclipsebazaar.com are synced from these cards by a drift-checked process. If your recompute disagrees with a card, that disagreement is itself a finding — the catalog's publish pipeline runs a claim gate on exactly that class of dispute.

What this corpus is not

Not a ranking, leaderboard, or security audit of third-party software: every card grades a Cyan Eclipse product, at a stated scope, against its own published specification. Scores and verdicts from different scopes are not comparable across products.

License

Compilation © 2026 Cyan Eclipse Technologies LLC. Published for verification, research, and citation; the underlying scorecards are the same files served from the Eclipse Bazaar catalog. Cite the card filename when you rely on a verdict.

---

Published by Cyan Eclipse Technologies LLC · cetech.ai · inquire@cetech.ai · Catalog: eclipsebazaar.com · Methodology: K-Check (full specification forthcoming)