Vendor: Cyan Eclipse Technologies · Platforms: MCP over stdio
Certification scope
CERTIFIED(local) — Eclipse Bazaar Test Bench 2026-10-25, 22 pass / 0 fail / 0 na of 22 checks. Run by W-003 (Verity), suites/blackbox.py, against bench-provisioned credentials and bench agent streams. Run 2 FAILED (U5: bare int() coercion in record_read's seq/limit filters escaping as a raw -32603) and run 1 PARTIAL cards retained append-only; the fix (coerce_int_param guard) was verified on the wire at raw JSON-RPC level, including a 114-probe schema-enumerated type fuzz over every parameter of all five tools, and the guard is ruled APPROVED as the house reference implementation for the raw-coercion defect class; run 3 is the standing card. U1 is a genuine pass (a real credential layer exists). N-check stands: the recorded agent cannot read, grade, or rewrite its own record; the recorder receipts its own gaps. Tail-truncation claim verified exactly as stated (detectable only against a previously exported pack head — the product claims no more).
What it does
- Tamper-evident flight recorder for agent actions: an append-only, hash-chained record store signed at ingest by a recorder-held key the recorded agent never touches
- Credential shape is the security model: agents hold per-agent submission credentials (append-only, own name only); the owner credential reads, verifies and exports but cannot append — the recorded agent cannot read, grade, or rewrite its own record; auth is not an oracle (unknown id and wrong secret refuse byte-identically)
- The recorder assigns sequence, timestamps, chain hashes and the evidence grade from proof state — E-SIGNED (live append), E-CHAINED (verified agent-side chain anchor), E-CLAIMED (imported assertion, rendered as a claim forever); a submitter-supplied grade is refused by name in every mode
- Skipped submission sequences are receipted by the recorder itself as signed INGEST_GAP events; claims about never-submitted actions surface in coverage as claimed-not-submitted next to the gap, and a claim contradicting a live record flags a conflict naming both sequences — winner unpicked
- Mandate refs are recorded as refs; a missing one renders NO_MANDATE_RECORDED, never an inference. There is no update, delete or redact verb at any tier, and export packs verify standalone offline
- Cluster 3 law binds: evidence only — no insurability verdicts, no safety verdicts, no root-cause conclusions (banned output classes quarantined at render). v0.1 is a local single-store recorder; the hosted multi-tenant recorder is owner-gated (G-C3-1) and not built
Tools
record_appendrecord_verify_chainrecord_readrecord_export_packrecord_status
Problems it addresses
- My agent's own logs are the only record of the incident — and it wrote them
- The agent skipped steps and backfilled the story; I need the gaps receipted by someone who isn't the agent
- Prove what the agent did — and what it never submitted — to a third party, offline
Not tested
- Hosted multi-tenant recorder and any CETech-hosted recording of customer data — owner gate G-C3-1 (spec §15); v0.1 is a local single-store recorder only.
- OS-level isolation on a shared host — v0.1 isolates by recorder-held key + credential scoping + an append-only engine with no mutation code path; it does not claim an OS boundary the build does not create.
- External chain anchoring (spec §6.3 v1 hardening) — not built; tail-truncation of the store file is therefore detectable only against a previously exported pack head, which the README states plainly.
- Concurrent writers — the store is a single JSON file grown per call; one recorder process at a time.
- Streaming ingestion adapters from third-party frameworks (spec §6.1 gated) and legal-hold export formats beyond the evidence pack (counsel review) — not in v0.1.
- Dogfood wiring (recording worker runs) — a later ops step per the build brief; this build ships the recorder plus its fixture stream only.
Bench scorecard: test-bench/results/eclipse-blackbox-mcp-2026-10-25-3.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 →