Fleet restart / recovery
Startup and recovery telemetry with Unknowns and explicit source units.
Public evaluation
Validation & Benchmarks
The demo suite is designed for blind diagnostic evaluation: the assistant sees evidence, not the planted answer, and the evaluator judges whether the reasoning is actually supported.
Five evidence domains
Different failure shapes, same evidence discipline.
The public suite spans more than 20,000 miner/evidence records across CSV, XLSX, JSONL, and NDJSON.
Network reachability
Shared segment evidence designed to separate unreachable devices from reachable zero-hash states.
Power & thermal
Topology, feed loading, and thermal evidence with competing infrastructure hypotheses.
Repair history
Longitudinal events for recurring-fault and repair-outcome reasoning.
Miner-log triage
Individual synthetic cases across healthy, startup, thermal, pool, PSU, hashboard, and other symptoms.
What a good run does
Accuracy is necessary, but not sufficient.
Extract faithfully
Preserve units, Unknown values, source meaning, and factual distinctions such as unreachable versus reachable zero-hash.
Reason transparently
Use useful derived metrics with visible denominators, recognize shared versus miner-specific patterns, and resist premature component diagnosis.
Reduce uncertainty
State confidence appropriately and finish with one discriminating read-only next check.
Current validation status
Claims should match evidence.
Synthetic demo suiteSIMULATED
Evidence methodologyTESTED FROM REAL EXPORT
Codex / Agent Skill packageSTRUCTURALLY VALIDATED
Claude pluginCLAUDE CLI VALIDATED
Physical miner integrationNOT PART OF v1.1.0
The repository intentionally does not publish hidden scoring answers or planted-case answer keys.
Blind use
Do not tell the assistant the planted pattern.
Use the public demo prompts, provide the evidence file, and judge whether the reasoning follows the supplied evidence rather than whether the assistant can restate an answer key.
Questions about validation, private pilots, or benchmark collaboration can be sent to austin@wnclogiclab.com.
View technical benchmark document →