Engineering that leaves
an evidence trail.
AlvenX is an independent open-source research and engineering practice. It turns a narrow technical question into a runnable system, preserves the result and its failures, and stops the claim where the evidence stops.
The practice
The first generation covers local multimodal measurement, browser-agent regression evidence, and calibration of physics-aware video metrics. These are separate projects, but they share one operating discipline: identify the exact system under test, use an independent oracle where possible, preserve negative evidence, and keep software maintenance separate from the research result.
A small set of rules, applied consistently.
Local ownership
Prefer workflows that can be run, inspected, and retained on hardware the user controls.
Evidence before claims
Separate measured facts, interpretation, assumptions, and unresolved questions.
Failures stay visible
Keep negative results, first failures, and stopped experiments useful to the next decision.
Observed problems authorize work.
Every new abstraction, dependency, module, interface, benchmark, or model integration must point to a concrete Issue, reproducible failure, user need, or registered hypothesis.
Software can improve without rewriting evidence.
The maintained projects are released at v1.1.2. Their published measurements and protocols remain attached to immutable v1.0.0 evidence.
One owner, explicit responsibility.
AlbertXXuu is the sole human code contributor for the first-generation repositories and owns the problem framing, protocol decisions, implementation acceptance, evidence boundaries, and release decisions. Automated assistance does not replace that responsibility.
First Generation is maintained, not expanded by default.
Correctness, security, portability, compatibility, and reproduction fixes remain welcome. New research or product work requires a separate evidence-backed decision.
Review the implementation and its evidence.
These cases connect a concrete problem to implementation choices, recorded verification, source code, and reproduction instructions. Each project states the boundary of its result so a technical reviewer can inspect the work directly.
Score the DOM before forced cleanup
A completed settings task was recorded as a failure because the browser closed before scoring. The correction preserves the page for an independent DOM check, while a separate checkout run shows why an agent's own success report cannot define the result.
Define timing and memory at observable boundaries
OpenMultimodalLab preserves 612 measured attempts across 102 tasks and two pinned models. The useful engineering result is a comparison that keeps task quality, compute timing, memory, interruption recovery, and evidence provenance distinct.
Keep the candidate readable and the outcome independent
A standalone Playwright candidate distinguished DrawDB revisions before and after Issue 687's fix. ReproLock preserved the repeated differential, build identity, and cleanup evidence, while the ordinary Playwright control exposed the limits of the product value claim.
Connect CPU fusion, native editing, and independent reimports
The table, two backed chairs and accepted stool form four editable roots with 37 closed parts. Original-frame owner-observation gates pass; the separately authored layout certifies all 507 cross-furniture pairs at at least 5 mm separation. Full-reference table reverse-area P95 remains 48.378 mm.
Results are useful only with their limits attached.
The First Generation closure report puts the core result, negative result, reproduction path, and stop decision for all three projects in one place.
Chinese typography uses MiSans, © Xiaomi Inc., distributed under the MiSans License Agreement. Latin text and numerals use Instrument Sans.