Your first audited inference

A complete worked example. Same shape as OpenAI Chat Completions, with a flat vetted block carrying the per-query environmental receipt.

The minimal call

from openai import OpenAI

client = OpenAI(
    base_url="https://api.vettedinference.com/v1",
    api_key=VETTED_API_KEY,
)

response = client.chat.completions.create(
    model="mistral-medium-3",
    messages=[
        {"role": "user", "content": "Explain CSRD double materiality in two sentences."}
    ],
)

print(response.choices[0].message.content)

What you get back

{
  "id": "chatcmpl-vi-9f3c2e",
  "object": "chat.completion",
  "created": 1746604320,
  "model": "mistral-medium-3",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "CSRD double materiality requires reporting on..."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": {
    "prompt_tokens": 24,
    "completion_tokens": 78,
    "total_tokens": 102
  },
  "routing_trace": {
    "route_policy": "lowest_carbon",
    "decision_mode": "degraded",
    "decision_reason": "degraded_no_operational_live_signal",
    "operational_signal": {
      "source": "energinet_forecast_live",
      "signal_kind": "forecast",
      "gco2_per_kwh": 46.0,
      "observed_at": "2026-05-19T12:00:00Z",
      "requested_at": "2026-05-19T12:05:00Z",
      "age_minutes": 5.0,
      "temporal_match": "live_window",
      "live_eligible": true
    },
    "receipt_signal": {
      "source": "ember",
      "signal_kind": "fallback",
      "gco2_per_kwh": 270.0,
      "observed_at": "2026-01-01T00:00:00Z",
      "requested_at": "2026-05-19T12:05:00Z",
      "age_minutes": 200736.0,
      "temporal_match": "annual_fallback",
      "live_eligible": false
    }
  },
  "vetted": {
    "query_id": "4c1d4f4d-7fb6-47f7-9d80-a9c2b149c7f6",
    "tier": "parametric",
    "gco2e": 0.21,
    "gco2e_ci90_lo": 0.15,
    "gco2e_ci90_hi": 0.31,
    "ml_water": 8.4,
    "ml_water_ci90_lo": 6.1,
    "ml_water_ci90_hi": 11.8,
    "methodology_version": "0.1.0",
    "methodology_hash": "abc123def4567890",
    "pedigree_score": 2.5,
    "route_taken": "scaleway:fr-par-1",
    "grid_zone": "fr-par-1",
    "grid_intensity_gco2_per_kwh": 58.0,
    "grid_intensity_source": "entsoe_derived_live",
    "grid_intensity_observed_at": "2026-05-17T12:00:00Z",
    "grid_intensity_requested_at": "2026-05-17T12:07:00Z",
    "grid_intensity_age_minutes": 7.0,
    "grid_temporal_match": "live_window",
    "audit_missing": false
  }
}

Reading the receipt

Field	Meaning
query_id	Stable ID for the receipt and audit-ledger record.
tier	One of proxy, parametric, telemetry, audited.
gco2e / gco2e_ci90_*	Best estimate and 90% interval for climate impact.
ml_water / ml_water_ci90_*	Best estimate and 90% interval for water impact.
pedigree_score	Composite data-quality score used by the methodology.
grid_intensity_gco2_per_kwh	Grid intensity value actually used in the calculation.
grid_intensity_source	Machine-readable provenance label, for example entsoe_derived_live, entsoe_derived_prev_week, neefe, ipcc_default, or native.
grid_intensity_observed_at	Timestamp of the source observation used.
grid_intensity_requested_at	Timestamp of the inference we were trying to match.
grid_intensity_age_minutes	Absolute difference between requested and observed timestamps.
grid_temporal_match	Match class such as live_window, prev_week_fallback, annual_fallback, or default_fallback.
audit_missing	true only when the response succeeded but the ledger write had to be queued for retry.

Reading the routing trace

The additive routing_trace block is operational, not receipt-grade. It tells you what the control plane used for routing and whether the request had to degrade out of strict carbon-aware routing.

Field	Meaning
decision_mode	normal when routing used a fresh operational live signal, degraded when it had to fall back to the standard provider-preference path.
decision_reason	Machine-readable reason such as operational_live_signal or degraded_no_operational_live_signal.
operational_signal.signal_kind	One of measured, forecast, provisional, or fallback.
receipt_signal.*	Conservative signal that actually fed the receipt calculation. This can be weaker than the operational signal.

Temporal integrity

For ENTSO-E receipts, entsoe_derived_live means the source timestamp was within +/-15 minutes of the inference timestamp. If that is not true, the system falls back and the receipt tells you so.

This is deliberate. We would rather return a clearly-labeled fallback than pretend an older source point was live.

Streaming

Streaming responses include the vetted block in the final SSE event:

data: {"choices":[{"delta":{"content":"..."}}]}
data: {"choices":[{"delta":{"content":"..."}}]}
...
data: {"vetted":{"query_id":"...","grid_intensity_source":"entsoe_derived_live",...}}
data: [DONE]