evals(community): OsaurusAI/ZAYA1-8B-MXFP4 on Apple M5 Max (partial, aborted)

[jax-0n-git]

Crowdsourced model-compatibility run (see COMMUNITY_EVALS.md).

Partial / aborted — not a full matrix. Do not treat 36/95 as a comparable leaderboard cell.

Why it was aborted

• Wall clock ~10.7 h on Apple M5 Max / 128GB before abort (contributor jax-0n-git).
• Initial 14-suite batch wrote almost nothing; loop fell back to serial per-suite retries.
• agentdb.queryable-saved-view hit the 1800s watchdog after a single step dumped ~47k reasoning chars and no tool call (stop=unknown).
• Model kept unbounded reasoning at ~14 tok/s; DefaultAgent was still on 24/49 after ~1.4 h on that suite alone with CPU ~120% (not a 0% hang).
• Remaining suites not run: DefaultAgent (incomplete), HTTPAPI, Memory, MicroPerf, PromptInjection, ReasoningChannel, Subagent, SandboxFrontier.

What is in this file

Completed suite JSONs only: AgentDB (watchdog-terminated), AgentLoop, AgentLoopFrontier, AppleScript, CacheProof, CapabilityClaims, ComputerUseLoop. Aggregated via osaurus-evals matrix + compat --validate. Self-judge.

Environment: Apple M5 Max · RAM 128GB · see contribution file.

One contribution file; no other changes.