Response Format Reference¶

Every response from the Geodesia G-1 gateway includes a geodesia extension object. This page is the authoritative reference for all fields in that object.

Gateway Response Structure¶

When you make a request to POST /v1/chat/completions, the response follows the standard OpenAI format with an additional geodesia key:

{
  "id": "chatcmpl-abc123",
  "object": "chat.completion",
  "created": 1749555000,
  "model": "mistralai/Mistral-7B-v0.3",
  "choices": [
    {
      "index": 0,
      "message": {
        "role": "assistant",
        "content": "The capital of France is Paris."
      },
      "finish_reason": "stop"
    }
  ],
  "usage": { ... },
  "geodesia": {
    "call_id": "call_abc123",
    "session_id": "sess_xyz",
    "prompt_blocked": false,
    "answer_blocked": false,
    "block_reason": null,
    "dominant_axis": null,
    "axes_available": ["halluc_context", "halluc_closedbook", "prompt_safety", "answer_safety", "jailbreak"],
    "scores": { ... },
    "thresholds": { ... },
    "watermark": { ... },
    "rag": null,
    "latency_ms": 312
  }
}

`geodesia` Object¶

Top-Level Fields¶

Field	Type	Description
`call_id`	`string`	Unique identifier for this call. Use this to look up the call in the audit chain or oversight queue.
`session_id`	`string`	Session grouping identifier.
`prompt_blocked`	`boolean`	`true` if the input prompt was blocked before generation. When `true`, `choices[0].message.content` is `null` (or contains a block notice).
`answer_blocked`	`boolean`	`true` if the generated answer was withheld due to a detection threshold.
`block_reason`	`string` \| `null`	Plain-language reason for blocking. `null` if not blocked.
`dominant_axis`	`string` \| `null`	Which detection axis was the primary reason for blocking when multiple axes triggered.
`axes_available`	`array[string]`	List of detection axes that were actually computed. Depends on which model is loaded and whether context was provided.
`scores`	`object`	Per-axis detection scores. See below.
`thresholds`	`object`	The threshold values that were applied (after any per-request overrides).
`watermark`	`object`	Watermark metadata. See below.
`rag`	`object` \| `null`	RAG retrieval results, if RAG was triggered. See below.
`latency_ms`	`integer`	Total wall-clock latency for this request in milliseconds.

`scores` Object¶

Contains one field per available detection axis:

"scores": {
  "prompt_safety": {
    "score": 0.12,
    "logit": -1.98,
    "triggered": false
  },
  "answer_safety": {
    "score": 0.08,
    "logit": -2.44,
    "triggered": false,
    "combined_score": 0.07,
    "combined_logit": -2.51,
    "combined_threshold": 0.11,
    "combined_triggered": false,
    "per_signal": { ... }
  },
  "halluc_context": {
    "score": 0.23,
    "logit": -1.19,
    "triggered": false,
    "combined_score": 0.18,
    "combined_logit": -1.52,
    "combined_threshold": 0.617,
    "combined_triggered": false,
    "per_signal": { ... },
    "n_signals": 10
  },
  "halluc_closedbook": {
    "score": 0.09,
    "triggered": false
  },
  "jailbreak": {
    "score": 0.03,
    "triggered": false
  }
}

Per-Axis Score Fields¶

Field	Description
`score`	Calibrated probability [0–1]. Use this for display and threshold comparison. Higher = more likely the issue is present.
`logit`	Raw logit (uncalibrated). Present for advanced users and debugging only.
`triggered`	`true` if `score` exceeded the configured threshold
`combined_score`	For `halluc_context` and `answer_safety`: the output of the calibrated multi-signal combiner, which aggregates several internal detection signals. This is the definitive score used for blocking decisions.
`combined_logit`	Raw logit from the combiner
`combined_threshold`	The threshold used for the combined score decision
`combined_triggered`	Whether the combined score triggered a block decision
`per_signal`	Breakdown of each internal signal's contribution to the combined score
`n_signals`	Number of signals included in the combined score

Which score to use

Always use combined_score when it is present (for halluc_context and answer_safety). The combined_score is calibrated with AUROC > 0.97 and is more reliable than the raw score. Use raw score for axes that do not have a combiner (halluc_closedbook, jailbreak, prompt_safety).

`thresholds` Object¶

The thresholds that were applied to produce the triggered values. These may differ from the default configuration if per-request threshold_overrides were provided.

"thresholds": {
  "prompt_safety": 0.35,
  "answer_safety": 0.11,
  "halluc_context": 0.617,
  "halluc_closedbook": 0.50,
  "jailbreak": 0.50
}

`watermark` Object¶

"watermark": {
  "token": "hmac:v1:a8b3c1d4e5f6...",
  "generated_by": "Geodesia G-1",
  "call_id": "call_abc123",
  "timestamp": "2026-06-10T10:23:45Z",
  "disclosure": "This content was generated by an AI system."
}

Field	Description
`token`	HMAC-SHA256 token for verification. Pass to `POST /v1/glad/watermark/verify` to confirm authorship.
`generated_by`	System identifier
`call_id`	Matches `geodesia.call_id`
`timestamp`	ISO 8601 generation timestamp
`disclosure`	Plain-language AI disclosure string (configurable in `watermark.disclosure_text`)

`rag` Object¶

Present when RAG retrieval was triggered for this request.

"rag": {
  "retrieved_chunks": [
    {
      "text": "Paris is the capital of France.",
      "source": "geography-kb/europe.pdf",
      "score": 0.97,
      "rank": 1
    }
  ],
  "collection_id": "my-knowledge-base",
  "query_used": "capital of France",
  "n_retrieved": 3,
  "reranked": true
}

Field	Description
`retrieved_chunks`	List of retrieved document chunks, sorted by relevance
`retrieved_chunks[].text`	The chunk text
`retrieved_chunks[].source`	Source document identifier
`retrieved_chunks[].score`	Retrieval relevance score (0–1)
`retrieved_chunks[].rank`	Rank position after reranking
`collection_id`	Which RAG collection was queried
`query_used`	The query sent to the vector store (may differ from the original prompt if the gateway rewrote it)
`n_retrieved`	Number of chunks retrieved
`reranked`	Whether the BGE reranker was applied

Blocked Response¶

When prompt_blocked or answer_blocked is true in blocking mode, the response content is replaced with a block notice:

{
  "choices": [
    {
      "message": {
        "role": "assistant",
        "content": "[Response withheld: answer_safety threshold exceeded (score: 0.78, threshold: 0.11)]"
      },
      "finish_reason": "content_filter"
    }
  ]
}

In passthrough mode, the original response is delivered and answer_blocked is set to true as an annotation.

Streaming Format¶

In streaming mode (stream: true), the geodesia object is delivered in the final chunk — the chunk with finish_reason: "stop" or "content_filter". All preceding chunks are standard OpenAI SSE format without the geodesia extension.

data: {"id":"chatcmpl-abc","choices":[{"delta":{"content":"Paris"},...}],...}
data: {"id":"chatcmpl-abc","choices":[{"delta":{"content":"."},...}],...}
data: {"id":"chatcmpl-abc","choices":[{"delta":{},"finish_reason":"stop"}],"geodesia":{...}}
data: [DONE]

Response Format Reference¶

Gateway Response Structure¶

geodesia Object¶

Top-Level Fields¶

scores Object¶

Per-Axis Score Fields¶

thresholds Object¶

watermark Object¶

rag Object¶