Two models, one verdict.

Cross-Architecture Audit checks whether two models agree on the same inputs — so you know which traffic is safe to migrate, distil, or push to the edge before you change anything in production.

Endpoint POST /api/v1/cross-model/probe • Dimensions can differ between models • Available on Compliance and Enterprise tiers

When to use it

Migration

Replacing a large model with a smaller one for cost or latency. See which inputs the two models still agree on before you switch.

Distillation

Student model trained from a teacher. Verify the student preserves the teacher’s behaviour on the inputs that matter most.

Edge deployment

Cloud model paired with an edge model. Audit per-input agreement so you know which traffic is safe to keep on-device.

What you provide

Embeddings from both models on the same set of inputs (in the same order), plus a class label per input. Both feature sets must cover identical observations — the audit reads behavioural agreement on a per-input basis.

features_a — float[n][d_a] from model A
features_b — float[n][d_b] from model B (d_a and d_b may differ)
labels — one label per input, shared between models
class_names — optional human-readable names

Call the API

curl

curl -X POST https://transferoracle.ai/api/v1/cross-model/probe \
  -H "X-API-Key: YOUR_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "features_a": [[0.12, -0.04, ...], ...],
    "features_b": [[0.31,  0.08, ...], ...],
    "labels":     ["dog", "cat", "bird", ...]
  }'

python

import numpy as np
import requests

# Embeddings from your two models on the SAME inputs, in the same order.
# Dimensions can differ (e.g. 1024 vs 768).
features_a = np.load("features_model_a.npy").tolist()
features_b = np.load("features_model_b.npy").tolist()
labels     = ["dog", "cat", "bird", ...]  # one label per input

resp = requests.post(
    "https://transferoracle.ai/api/v1/cross-model/probe",
    headers={"X-API-Key": "YOUR_KEY"},
    json={
        "features_a": features_a,
        "features_b": features_b,
        "labels": labels,
    },
)
print(resp.json())

What you get back

{
  "verdict": "DEGRADED",
  "compatibility_score": 0.74,
  "coverage": 0.91,
  "n_classes": 10,
  "n_samples": 5000,
  "per_class": [
    { "label": "dog",  "n_samples": 500, "agreement": 0.92 },
    { "label": "cat",  "n_samples": 500, "agreement": 0.88 },
    { "label": "bird", "n_samples": 500, "agreement": 0.41 },
    "..."
  ],
  "classes_at_risk": ["bird", "frog"],
  "recommendations": [
    {
      "priority": "medium",
      "action": "Validate the listed classes on representative deployment data before switching models.",
      "evidence": "2 classes show partial behavioural divergence."
    }
  ]
}

SAFE≥ 0.85

Models agree on the great majority of inputs. Migration is low-risk.

DEGRADED0.65 – 0.85

Partial divergence. Review the listed classes before switching production traffic.

CRITICAL< 0.65

Behavioural divergence is severe. Do not migrate without per-class validation.

compatibility_score is the overall fraction of inputs where the two models agree. coverage is the fraction of inputs the audit could analyse. classes_at_risk lists labels where per-class agreement falls below the DEGRADED threshold — these are the inputs to validate before migration.

Get started

Cross-Architecture Audit is available on Compliance and Enterprise tiers. Claim a free API key first, then upgrade if your tier doesn't include this endpoint.

Get an API Key View Plans