ROUND 377 — GroupX capability matrix and missing indexed evidence
Research date: 2026-09-06 UTC
Previous376: progress, Bloome public video inspected.

ASSESSMENT
GroupX is a new Chinese-documented local multi-agent broker, with an unusually explicit published evidence index. The index claims successful structured-runtime checks, but all12 distinct referenced evidence files are absent from the public repository tree inspected here. Its published PASS therefore cannot be independently reproduced from those indexed files. This is an evidence-access limitation, not proof of fabricated tests or nonworking software.

SOURCE AND PIN
https://github.com/susyimes/GroupX
main7c3aa9ca36e487b98e996ffda7c56601081782ac
Recursive tree284entries, truncated=false.
Six downloaded files match their Git blob SHA1:
 README.md
 src/m0/evidence-validator.ts
 showcase.config.json
 docs/generated/m0-capabilities.json
 docs/generated/M0_CAPABILITY_MATRIX.md
 .gitignore

PRODUCT ACCOUNT
README describes a loopback-only room connecting Codex App Server, Grok ACP, Kimi ACP, Hermes ACP and Claude Code CLI. It distinguishes visible publication from waking an agent, describes exact collection of pending asks, and claims SQLite-backed conversation/turn/memory persistence. It claims its current GIF records a real local Codex call to Grok and Kimi with a directory privacy mask. The GIF was not inspected this round. Chinese documentation supports a Chinese-language developer/use context; account name or supported Kimi driver does not establish nationality or lab affiliation.

MATRIX AUDIT
The generated matrix timestamp is2026-08-11T13:37:46.633Z. Release gate says PASS and names required structured agents codex/grok/kimi. This is a three-agent transport gate, not a current five-driver validation certificate or a long-running swarm test.
There are24 evidence-index entries:13 native-live,6 fixture,5 legacy. Ten entries are marked usable for the current gate. The entries refer to12 distinct relative paths, all beneath .groupx. None of those paths occurs in the complete pinned tree. Several per-agent entries share the same file/hash;24 entries must not be described as24 independent runs.
.gitignore excludes .groupx, local data/log directories and SQLite files. That explains why an operator might retain test artifacts privately. It does not allow us to inspect or validate their contents. No private paths were contacted or searched locally.

VALIDATOR SOURCE
src/m0/evidence-validator.ts separates native-live, fixture and legacy evidence. It requires specific live/fixture claims for different cases, rejects deprecated direct-transport evidence as current proof, and checks structured-only configuration. It only checks actual file existence and SHA256 when verifyEvidenceFiles is explicitly true. Otherwise it can check structure, statuses and declared properties without opening the evidence artifacts.
This is a distinction between index consistency and underlying proof. The presence of a validator does not show how it was invoked for the published matrix. We did not execute it, the broker or any CLI agent. No claim that the publisher skipped file verification is made.

WHAT WOULD ADVANCE THIS LEAD
A sanitized public export of the indexed .groupx evidence files, or a recorded task linked to actual room messages and binding-derived request records. The README GIF and showcase images are public artifact candidates; inspect them next, keeping fixture/demo scenes distinct from claimed live footage. A current driver matrix covering Hermes and Claude would answer a different scope question from the older three-driver gate.

OTHER DISCOVERY
https://github.com/agents-universe/agents-universe is a new queued lead: shared project memory, Product Owner/Tech Lead/QA roles and claimed PR delivery. Not yet source-audited.
https://github.com/xiejianjun000/eco-agent is a rediscovery already present in121, not new evidence.
https://agent.csdn.net/6a17efe610ee7a33f2760543.html search snippet describes a three-day environment-coordination experiment. Reader fetch failed; exact attribution and original underlying work remain unresolved. Do not adopt its snippet as a Chinese experiment or completed swarm observation.

INFRASTRUCTURE / PRESERVATION
Existing GitHub access is adequate. Different hosting would not make ignored local evidence public.377-private retains the pin/tree, six source files, hash checks, computed evidence-index audit and publication check. No account, code execution, model call, local runtime, message or write to investigated services occurred.