Introducing Guardal Holdfast
Holdfast is sensor intelligence that fuses distributed tracks into one current picture on the sensors an organization already owns under human release authority.
Guardal is releasing Holdfast as software that runs above the radar and electro-optical and passive RF sensors a force already fields and above the effectors already on its rails. What Guardal builds is the intelligence between them.
Holdfast fuses what those sensors return into one track picture and keeps the uncertainty and the source attached to every track in it. When a hundred cheap aircraft arrive over the same volume at once that picture is what an operator works from.
Holdfast allocates effectors against the picture and replans while the raid is still changing. A human holds release authority on every engagement.
Jamming and denied positioning are the conditions Holdfast is designed for. Each node keeps deciding on what it can still see inside the authority delegated to it and the shared picture reconciles once the network returns.
Integration runs through the interfaces a program already exposes. Its sensors and its radios and its message formats and its command systems and its effectors are the surface Holdfast meets and mixed vendors across mixed generations are the ordinary case. The fusion and allocation internals stay unpublished. Every figure below is a hardware-in-the-loop measurement and each one states the workload it was taken at.
Holdfast is licensed per protected site and it is available for technical demonstrations and program-scoped pilots with government and defense and advanced research organizations.
What Holdfast is measured at
Every requirement and every measured result with its workload and its environment and the scalability ceiling and the degradation curve. Each figure states the conditions that make it readable.
Where the numbers are taken
Workload
Tracks and agents injected at a stated rate. The decision figure was taken at 64 agents and 2,500 active tracks.
Nodes under test
The nodes under test are real hardware running the real build. Sustained track handling reaches 5,000 tracks sustained.
Injected loss
Packet loss applied to the links between nodes. Allocation quality holds 94% at 25% loss.
Decision tap
Read where a track becomes a decision. 180 ms p95.
Allocation tap
Read where a decision becomes an assignment. 95 ms p95.
Command tap
Read at the far end where sensor input has become command output. 650 ms p95.
Timebase
One clock over the whole bench. Replanning holds 10 Hz sustained.
Design requirements
Decision latency
Under 500 ms
A quadcopter closing at 30 m/s covers 15 m while the loop thinks. The bar is set by how much ground the defender can afford to give up per decision.
Task allocation latency
Under 250 ms
Allocation runs many times per engagement as tracks split and effectors commit. It has to cost a fraction of the decision budget.
Replanning rate
10–20 Hz
A swarm replans around losses continuously. A defense that replans slower than the attacker adapts is solving a picture that has already moved.
Simultaneous tracks
1,000–10,000
Mass is the attack and the picture has to hold the whole raid. Measured separately at 5,000 tracks sustained and reported below.
Simultaneous engagements
50–200+
Engagements in flight at once is what separates coordinated defense from a queue that serves one threat at a time.
Packet-loss tolerance
Allocation quality
Jamming degrades a link partially and asymmetrically. The requirement is graceful degradation across the whole range with a figure stated at every point on it.
Operator ratio
Operator supervision
A defense that adds a person per effector runs out of people. A human releases every engagement and the ratio is what decides whether that authority stays affordable.
E2E Latency
Under 1 s
The sum of every stage the loop passes through. It is the only figure a defended asset actually experiences.
Measured results
Decision latency
180 ms p95
Task allocation latency
95 ms p95
End-to-end latency
650 ms p95
Replanning rate
10 Hz sustained
Decision latency
180 ms p95
64 agents and 2,500 active tracks · Hardware-in-the-loop
Task allocation latency
95 ms p95
64 agents under a dynamic reassignment workload · Hardware-in-the-loop
Replanning rate
10 Hz sustained
Contested communications scenario · Hardware-in-the-loop
End-to-end latency
650 ms p95
Sensor input to decision to command output · Hardware-in-the-loop
Scalability result
A separate run past the benchmark operating point. Every latency figure above belongs to its own operating point.
Sustained track handling
5,000 tracks sustained
Dedicated scalability testing with multi-source sensor fusion · Hardware-in-the-loop
Allocation quality against packet loss
Quality holds nearly flat through 25% loss and gives up seven points by 50% and falls away past that.
0% packet loss · 98% allocation quality retained
Media inquiries · media@guardal.ai
