Did the work succeed?
Measures task completion, factual correctness, and whether the output meets the stated operating requirement.
Benchmark Registry
Score the system, not the demo.
Benchmarks are the public registry for how ALCUB3 evaluates runtime quality, trust boundaries, field signals, and failure modes before those claims become product or mission language.
Platform
Runtime quality
Task success, tool correctness, memory continuity, and approval-path behavior.
Impact
Signal quality
Coverage, calibration, confidence, and source freshness for public-interest work.
Status
Registry live
Formal result tables publish as each benchmark clears review.
Current benchmark families
Each benchmark has a claim, method, and destination.
| Benchmark | Status | What it measures | Destination |
|---|---|---|---|
| Runtime inspection | In Review | Task trajectory quality, tool-call correctness, recovery, and approval-ready output. | Platform track |
| Control-plane failure modes | Evidence Gathering | Privilege routing, instruction conflict handling, privacy-aware coordination, and reliability under review. | Reports |
| Memory-backed continuity | Planned | How well workspace memory preserves useful context without leaking authority or stale assumptions. | Learning Hub |
| Water intelligence calibration | Planned | Coverage, regional fit, score-vs-ground-truth, and evidence-tier correctness. | Impact track |
Evaluation rule
Failure modes are part of the result.
How did it get there?
Captures unnecessary steps, risky tool calls, prompt drift, missing escalation, and whether the reasoning path stayed legible.
Did authority stay bounded?
Checks identity, permission, policy, approval, and recovery boundaries before calling the result production-ready.
Cross-links
Benchmarks feed publications, product pages, and lessons.
Methods
Every benchmark references the method page that defines scoring, evidence tiers, and caveats.
Read methods →Reports
Benchmark results become reports when they produce a meaningful product, market, or mission conclusion.
Open reports →Learning
Benchmark logic becomes learning content when operators or builders need to use the pattern themselves.
Open Learning Hub →