Start with the specimen.
The product treats the contract as the object of investigation. Runtime bytecode, read-only probes, proxy storage, and risk-surface checks establish the material the system can reason about.
Investigation system · Nuvixes
Crash Lab
LiveA smart-contract investigation workspace that connects read-only observations to evidence-cited AI explanations.
Why this exists
A confident explanation is not evidence.
When a smart contract looks suspicious, an impressive answer can be dangerously comforting. Crash Lab starts with deployed code and asks the explanation to stay attached to what can actually be inspected.
Make the explanation accountable.
The AI layer selects complementary read-only probes and cites evidence IDs in its findings. Unsupported citations are discarded. The aim is an investigation someone can follow, rather than a verdict whose confidence has to be taken on trust.
The build had to work outside the original machine.
The August release work repaired clean installation on Linux, removed unreachable legacy helpers, and documented the DeepSeek credit release gate. Those changes are less visible than the industrial visual identity, but essential to making the demonstration repeatable.
Build journal
The turns
along the way.
Recorded milestones from the build. Dates describe these events, not the entire project history.
Clean installation repaired
The repository records a Linux npm clean-install repair.
Release boundaries documented
Legacy helpers were removed and the model-credit release gate was documented.
Where it stands
What works today.
A published contract-investigation prototype with an evidence-cited AI workflow. Its demo explicitly avoids treating read-only observations as proven exploits.
The next chapter
What comes next.
Improve evidence coverage and make uncertainty easier to inspect alongside each finding.
Based on the original Crash Lab demo and August 15 repository history.
Continue exploringDownrail