Model Profile Registry

The Model Profile Registry is a Primust-maintained signed registry of empirical per-operator drift profiles for governance-relevant ML models. It's what activates Bounded Inference — the proof level for HuggingFace transformers.

What it is

Each profile records how much a model's per-operator outputs drift across different hardware classes (CPU, A10G, A100, H100). Drift is measured over 1000 calibration passes at the 99th, 99.9th, and 99.99th percentile for each operator type (Linear/MatMul, LayerNorm, Softmax, GELU, Dropout).

Profiles are signed by Primust's GCP KMS key and verifiable offline against Primust's public key. A tampered profile is detectable the same way a tampered VPEC is detectable.

When primust verify sees a bound_committed_inference VPEC, it resolves the profile_id to the profile and checks whether the committed merkle_root is consistent with the declared model class and GPU class. This is the additional verification step that makes Bounded Inference stronger than Execution.

Profile format

{
  "profile_id": "primust/distilbert-class/v1.2.0",
  "model_class": "distilbert",
  "architecture": "DistilBertForSequenceClassification",
  "calibrated_gpu_classes": ["cpu", "a10g", "a100", "h100"],
  "calibration_passes": 1000,
  "calibration_date": "2026-03-17",
  "safety_margin": 2.0,
  "operators": {
    "Linear":    { "p99": 2.86e-4, "p99_9": 5.49e-4, "p99_99": 7.93e-4 },
    "LayerNorm": { "p99": 3.19e-5, "p99_9": 3.91e-5, "p99_99": 4.23e-5 },
    "Softmax":   { "p99": 4.59e-6, "p99_9": 5.78e-6, "p99_99": 7.55e-6 },
    "GELU":      { "p99": 3.43e-5, "p99_9": 4.20e-5, "p99_99": 4.96e-5 },
    "Dropout":   { "p99": 4.88e-4, "p99_9": 7.32e-4, "p99_99": 9.52e-4 }
  },
  "profile_signature": "Ed25519:...",
  "signed_by": "primust.com/.well-known/primust-pubkey.pem"
}

The safety_margin: 2.0 means the thresholds are twice what was empirically measured. A DistilBERT-class profile covers both distilbert-base-uncased and distilbert-base-multilingual-cased — empirical calibration showed GELU and LayerNorm drift transfers within 1.25x ratio across both models, so a 2x safety margin provides full coverage.

How SDK lookup works

1. SDK extracts onnx_model_hash from the model at manifest registration or first p.record_check() call
2. SDK calls GET /api/v1/registry/lookup?hash={onnx_model_hash}
3. If profile found: downloads and caches locally; sets stage_type: bound_committed_inference
4. If not found: sets stage_type: open_source_ml (Execution fallback); queues model_profile_missing advisory gap
5. Cached profile is used for all subsequent inferences — no network call at inference time

# What the SDK emits when no profile is found:
# ⚠ model_profile_missing (Medium)
#   No Primust operator-bound profile found for model sha256:a4f9...
#   This check is issuing at Execution level.
#   To upgrade to Bounded Inference, request profile calibration at
#   app.primust.com/policy/registry

Supported models

The initial registry ships with a DistilBERT-class profile covering the most common governance classifiers.

Primust does NOT host these models. You download from HuggingFace normally. The registry contains only calibration profiles — metadata about how the models behave across hardware, not the model weights themselves.

Request calibration for your model

If your model isn't in the registry, submit a calibration request. Primust calibrates the drift profile, publishes it, and the model_profile_missing gap auto-resolves.

Via dashboard

1. Go to app.primust.com/policy/registry
2. Models detected in your manifests appear with status ✅ or ⚠
3. Click "Request calibration" for models showing ⚠
4. Optionally provide HuggingFace model ID to speed up identification
5. Select desired GPU classes
6. You'll receive an email when the profile is published (SLA: 7 days standard, 48h Enterprise)

Via API

curl -X POST https://api.primust.com/api/v1/registry/calibration-requests \
  -H "Authorization: Bearer pk_live_xxx" \
  -H "Content-Type: application/json" \
  -d '{
    "onnx_model_hash": "sha256:...",
    "huggingface_model_id": "my-org/my-classifier",
    "desired_gpu_classes": ["cpu", "a10g", "a100"]
  }'

# Response:
{
  "request_id": "cal_req_abc123",
  "status": "queued",
  "estimated_completion": "7d",
  "notify_email": "admin@your-org.com"
}

Verify a profile

Profiles are offline-verifiable against Primust's public key, the same way VPECs are.

# Download the profile
curl https://api.primust.com/api/v1/registry/profiles?model_hash=class_profile_distilbert \
  -H "Authorization: Bearer pk_sb_xxx" > profile.json

# Verify the profile signature offline
curl https://primust.com/.well-known/primust-pubkey.pem > primust-pubkey.pem
primust verify-profile profile.json --trust-root primust-pubkey.pem

# Output:
# ✓ Profile signature valid (Ed25519, primust.com)
# Profile: primust/distilbert-class/v1.2.0
# Calibration: 2026-03-17 · 1000 passes · A10G + A100 + H100 + CPU

API reference

List profiles

GET /api/v1/registry/profiles
Authorization: Bearer pk_live_xxx

# Response:
{
  "profiles": [
    {
      "profile_id": "primust/distilbert-class/v1.2.0",
      "model_class": "distilbert",
      "architecture": "DistilBertForSequenceClassification",
      "calibrated_gpu_classes": ["cpu", "a10g", "a100", "h100"],
      "calibration_date": "2026-03-17",
      "safety_margin": 2.0,
      "profile_signature": "Ed25519:...",
      "signed_by": "primust.com/.well-known/primust-pubkey.pem"
    }
  ]
}
# Note: operators{} thresholds omitted in list view. Include ?model_hash= to get full detail.

Lookup by model hash

GET /api/v1/registry/lookup?hash={onnx_model_hash}
Authorization: Bearer pk_live_xxx

# 200 Found:
{
  "found": true,
  "profile_id": "primust/distilbert-class/v1.2.0",
  "model_class": "distilbert",
  "architecture": "DistilBertForSequenceClassification",
  "profile_signature": "Ed25519:...",
  "signed_by": "primust.com/.well-known/primust-pubkey.pem"
}

# 404 Not found:
{ "found": false }

Request calibration

POST /api/v1/registry/calibration-requests
Authorization: Bearer pk_live_xxx
Content-Type: application/json

{
  "onnx_model_hash": "sha256:...",           // required
  "model_description": "Internal bias classifier for loan decisions",
  "huggingface_model_id": "my-org/my-model", // optional — helps Primust identify model
  "desired_gpu_classes": ["cpu", "a10g"]     // which GPU classes to calibrate for
}

# Response:
{
  "request_id": "cal_req_abc123",
  "status": "queued",
  "estimated_completion": "7d",    // "48h" for Enterprise tier
  "notify_email": "admin@your-org.com"
}