sentinel-01-pub-ths

sentinel-01-pub-ths is a TorchScript packaging of the public sentinel-01-pub Aurelex Sentinel Stage A artifact. It is provided for users who want to avoid Hugging Face remote custom Python code while evaluating the same limited-functionality wealth-management communications risk model.

This is not a production Aurelex model and must not be treated as legal, compliance, or investment advice.

Publisher And Ownership

• Model developed by Aurelex AI Corp.
• Published in collaboration with Ratio1.
• Contact: hello@aurelexai.com.
• All intellectual property rights in the model remain with Aurelex AI Corp.

This repository is intended to publish only the designated limited-functionality graph artifact, tokenizer files, schema, and evaluation metadata. It does not include proprietary training data, system prompts, production models, or internal Aurelex architecture details beyond the information needed to load and evaluate this public artifact.

Identity

• Repo ID: AurelexAI/sentinel-01-pub-ths
• Source public repo: AurelexAI/sentinel-01-pub
• Model key: sentinel-mb-c-d11
• Current model artifact version: sentinel-mb-c-d11-20260424
• Release channel: sentinel-01-pub-ths
• Source encoder lineage: answerdotai/ModernBERT-base
• Artifact format: torchscript_graph
• Publication status: public, approved by Aurelex on 2026-04-28

The graph artifact is exported from the same self-contained end-to-end public model served by AurelexAI/sentinel-01-pub. The source encoder lineage is listed for transparency and Hugging Face discoverability; users do not need to load a separate encoder.

Loading Without Remote Custom Code

Install the runtime dependencies:

pip install torch transformers huggingface_hub

Run the graph directly with PyTorch and the published tokenizer/schema:

import json
import torch
from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer

REPO_ID = "AurelexAI/sentinel-01-pub-ths"

model_path = hf_hub_download(REPO_ID, "model.torchscript.pt")
schema_path = hf_hub_download(REPO_ID, "sentinel_schema.json")

with open(schema_path, "r", encoding="utf-8") as handle:
  schema = json.load(handle)

tokenizer = AutoTokenizer.from_pretrained(REPO_ID, trust_remote_code=False)
model = torch.jit.load(model_path, map_location="cpu")
model.eval()

text = (
  "Subject: Portfolio review follow-up. Hi Karen, following our quarterly "
  "review, I recommend trimming part of the concentrated technology position "
  "and reallocating the proceeds into the municipal bond ladder we discussed."
)

inputs = tokenizer(
  text,
  return_tensors="pt",
  truncation=True,
  max_length=schema.get("max_length", 512),
)

with torch.no_grad():
  raw_outputs = model(inputs["input_ids"], inputs["attention_mask"])

logits = dict(zip(schema["output_order"], raw_outputs))

Decode the graph logits with the published label schema:

def decode_head(name, values):
  spec = schema["output_signature"][name]
  kind = spec["type"]

  if kind == "binary":
    probability = float(torch.sigmoid(values)[0])
    threshold = schema["thresholds"].get(name, 0.5)
    return {
      "label": probability >= threshold,
      "probability": probability,
      "threshold": threshold,
    }

  if kind == "multiclass":
    labels = spec["labels"]
    probabilities = torch.softmax(values, dim=-1)[0]
    index = int(torch.argmax(probabilities))
    return {
      "label": labels[index],
      "probability": float(probabilities[index]),
      "probabilities": {
        label: float(probabilities[position])
        for position, label in enumerate(labels)
      },
    }

  if kind == "multilabel":
    labels = spec["labels"]
    probabilities = torch.sigmoid(values)[0]
    threshold = schema["thresholds"].get(name, 0.5)
    return {
      "labels": [
        label
        for position, label in enumerate(labels)
        if float(probabilities[position]) >= threshold
      ],
      "threshold": threshold,
      "probabilities": {
        label: float(probabilities[position])
        for position, label in enumerate(labels)
      },
    }

  raise ValueError(f"unsupported output head type: {kind}")


result = {
  name: decode_head(name, values)
  for name, values in logits.items()
}

print(result)
print(schema["model_version"])

This loading path does not use Hugging Face remote custom Python code. The tokenizer is loaded with trust_remote_code=False, and the model graph is loaded explicitly with torch.jit.load.

Important: a TorchScript file is still an executable model artifact. Only load it from repositories you trust and pin a reviewed Hub revision for reproducible deployments.

Outputs

The decoded output matches the sentinel-01-pub Sentinel Stage A contract: violation, severity, domain, subtype, jurisdiction, why, impacted_principles, remediation_actions, content_type, audience_segment, detection_difficulty, and aggravating_factors.

These outputs are risk-review signals for human review. They are not final compliance determinations.

Evaluation

Dataset: 2026-04-07-final-audit-clear-v1, test split size 150.

Metric	Test
Stage-A	0.751
Violation F1	0.993
Severity Acc	0.727
Domain F1	0.803
Subtype F1	0.738
Jurisdiction Acc	0.740
Why F1	0.684
Principles F1	0.703
Remediation F1	0.618
Aggravating F1	0.655

Repository Contents

• model.torchscript.pt: TorchScript graph artifact containing the public end-to-end model.
• sentinel_schema.json: output head order, label schema, thresholds, model version, and runtime metadata.
• tokenizer files: tokenizer assets used by the graph.
• metadata.json: dataset signature, output signature, thresholds, and release metadata.
• metrics.json: evaluation metrics for the selected model.
• results.md: human-readable evaluation artifact.
• checksums.sha256: SHA-256 checksums for the published graph package files.

Intended Use And Limits

This model is intended for public demonstration and evaluation of automated first-pass risk signals in wealth-management communications. It is scoped to English client-communications examples under the dataset contract listed above.

Do not use this model as a legal decision-maker, a substitute for qualified compliance review, a general-purpose moderation system, or evidence of performance outside the stated dataset scope. Aurelex AI Corp may request modification or removal of this repository at any time.