Opening an ECU file in a hex editor is easy. Understanding it is not. Which bytes represent code, calibration data, checksums, or padding? Is the file a full read or one memory region? A pattern that looks like a map may still be a coincidence.
Watcher Eye starts from that uncertainty. The goal is not to ask one language model for a confident answer. It is to combine deterministic tools, specialized analysis agents, and a human reviewer who owns the final decision.
Separate facts from hypotheses
Every finding should carry a type:
- a fact produced by a deterministic parser or checksum routine;
- a hypothesis supported by patterns and references;
- an unresolved region that needs more data;
- a human decision with a recorded reason.
This distinction matters more than a polished interface because a plausible automotive claim can still be unsafe.
Build an evidence graph
Store offsets, byte ranges, endian assumptions, decoded values, tool versions, and parent findings. An agent should cite these records rather than produce an isolated paragraph. Conflicting interpretations remain visible instead of one silently replacing another.
Give each agent a narrow role
A useful pipeline can include agents for file identification, entropy and segmentation, processor signatures, calibration candidates, checksum families, external reference matching, and contradiction review. Agents exchange structured findings, not free-form conclusions.
{
"kind": "hypothesis",
"region": { "start": 131072, "end": 135168 },
"claim": "Possible calibration table",
"confidence": 0.62,
"basis": ["smooth numeric surface", "axis-like boundary pattern"]
}Keep safety boundaries explicit
Analysis is not authorization to modify or flash a control unit. The product should distinguish read-only exploration, simulated transformation, exported patch, and hardware action. High-impact steps require compatible hardware information, backups, checksums, and human approval.
Make uncertainty useful
The system should say what additional evidence would change its conclusion: a known-good file, a processor ID, an A2L definition, a second software version, or a physical measurement. That turns uncertainty into the next investigation step.
Watcher Eye is valuable when it helps an expert reason faster without hiding ambiguity. The product should be auditable, conservative around safety, and comfortable saying “unknown.”