TRIDENT v1 - Windows Process Security Classifier
Fine-tuned ONNX model that scores Windows process-start events and returns two verdicts: an action (allow / terminate) and a trust level (system_critical / known_safe / unknown / known_unsafe).
This is the decision kernel for TRIDENT, the immune layer for Windows.
Files
- trident-v1.onnx - 565 KB, the ONNX graph
- trident-v1-attn4.onnx.data - 276 MB, external weights. Filename must not change.
- tokenizer.json - tokenizer vocabulary
- tokenizer_config.json - tokenizer config
- rl_agent_config.json - inference config
Important: the ONNX graph references its weights as trident-v1-attn4.onnx.data. Both files must be in the same directory at load time, and the .data file must keep its exact name.
Inputs
- input_ids: int64 [B, seq_len]
- attention_mask: int64 [B, seq_len]
- marker_pos: int64 [B, num_options]
- marker_mask: bool [B, num_options]
- qtype: int64 [B]
Outputs
- logits: float32 [B, num_options]
Two questions per process
- Q1 action (2 options): 0=allow, 1=terminate
- Q2 trust_level (4 options): 0=system_critical, 1=known_safe, 2=unknown, 3=known_unsafe
Escalate is not a model class - it is a policy decision made by the TRIDENT cascade.
Intended use
Designed to run inside the TRIDENT pipeline, where its output is routed through a three-tier confidence cascade and a system-path safety net before any action is taken.
Direct use is not recommended. The model has a measured ~12 percent false-positive rate on core Windows binaries (csrss.exe, lsass.exe, winlogon.exe, powershell.exe). The TRIDENT cascade overrides these before any process is terminated.
Limitations
- Windows-only training data.
- Over-flags powershell.exe and executables in user-writable directories.
- Confidence is not temperature-calibrated.
- Scores process start only - no CPU, network, or file-write visibility.
License
Apache 2.0. See LICENSE-MODEL in the TRIDENT repository.