JSON Report Contract Proposal

Published Confluence page for Project Palisade.

Confluence page ID: 6104645647 Parent folder ID: 6018662571 Remote version: 1 Last remote update: 2026-07-01T13:46:21.984Z Sync status: Published to Confluence.

Purpose

This document proposes the first JSON report shape for Palisade guardrail results.

The report is the primary contract returned by Palisade. A remediated output may be returned as a separate payload or artifact reference, but the report explains whether the input was compliant, what domains were evaluated, what was detected, and what changed during remediation.

This is a concrete example for discussion, not a final strict API schema.

Contract Principles

• The first field shown in examples is compliant.
• compliant: true means the evaluated input/output is acceptable under the selected policy profile.
• compliant: false means at least one required domain failed or the final remediation did not pass policy.
• verdict uses the existing Palisade vocabulary: allow, flag, remediate, or block.
• Domain results should be structured enough for deterministic validation and LLM-as-judge evaluation.
• Evidence should identify what was detected without exposing unnecessary sensitive content.
• Remediation details should explain what was fine, what changed, and whether the remediated output passed the follow-up check.

Example Report

{
  "compliant": false,
  "report_id": "report_01HV7PalisadeExample",
  "schema_version": "0.1.0-proposal",
  "execution_id": "exec_2026_07_01_000123",
  "conversation_id": "conv_optional_456",
  "created_at": "2026-07-01T12:00:00Z",
  "application": {
    "app_id": "libre-assist",
    "environment": "pilot"
  },
  "policy_profile": {
    "profile_id": "libre-assist-base-v0",
    "profile_version": "0.1.0",
    "mode": "shadow"
  },
  "input": {
    "modality": "text",
    "content_ref": "request.input[0]",
    "normalized_artifacts": [
      {
        "artifact_id": "norm_text_001",
        "type": "normalized_text",
        "returned_to_caller": false
      }
    ]
  },
  "verdict": "remediate",
  "summary": "The request is mostly acceptable, but one domain requires redaction before the content can be returned.",
  "domains": {
    "violence": {
      "compliant": true,
      "detected": false,
      "severity": "none",
      "reason": "No violent threat, instruction, or graphic violent content was detected.",
      "evidence": [],
      "action": "allow"
    },
    "pii": {
      "compliant": false,
      "detected": true,
      "severity": "medium",
      "reason": "The input contains a direct email address that should not be retained in the final response.",
      "evidence": [
        {
          "type": "entity",
          "label": "EMAIL",
          "location": "input.text[42:61]",
          "safe_excerpt": "{EMAIL}"
        }
      ],
      "action": "remediate"
    },
    "medical_advice": {
      "compliant": true,
      "detected": false,
      "severity": "none",
      "reason": "No diagnosis, treatment change, dosing change, or clinical instruction was detected.",
      "evidence": [],
      "action": "allow"
    }
  },
  "remediation": {
    "eligible": true,
    "applied": true,
    "strategy": "text_redaction",
    "summary": "The email address was replaced with a neutral placeholder.",
    "what_was_fine": [
      "The request intent was within the allowed product-support scope.",
      "No unsafe medical instruction was detected."
    ],
    "what_changed": [
      {
        "domain": "pii",
        "change_type": "redaction",
        "before": "{EMAIL}",
        "after": "[redacted email]"
      }
    ],
    "output_ref": "remediation.output.text",
    "post_remediation": {
      "checked": true,
      "compliant": true,
      "verdict": "allow",
      "reason": "The remediated output no longer contains the detected email address."
    }
  },
  "module_results": [
    {
      "module_id": "pii-redaction-v0",
      "module_type": "deterministic",
      "version": "0.1.0",
      "domains": ["pii"],
      "verdict": "remediate",
      "confidence": 0.98
    }
  ],
  "telemetry": {
    "latency_ms": 184,
    "model_calls": 0,
    "policy_cache_hit": true
  }
}

Domain Result Shape

Each evaluated domain should use the same basic fields:

Field	Purpose
compliant	Domain-level pass/fail under the selected policy
detected	Whether the module detected the domain at all
severity	Suggested values: none, low, medium, high, critical
reason	Short human-readable explanation
evidence	Bounded evidence, offsets, labels, references, or safe excerpts
action	Suggested action: allow, flag, remediate, or block

Remediation Result Shape

The remediation object should explain both the decision and the final result:

Field	Purpose
eligible	Whether policy permits remediation for this finding
applied	Whether remediation actually ran
strategy	Strategy used, such as text_redaction, image_blur, or audio_mute
summary	One-sentence summary of the remediation
what_was_fine	Parts of the content that did not need changes
what_changed	Specific changes made, using safe placeholders when needed
output_ref	Reference to the separate remediated output payload or artifact
post_remediation	Follow-up safety check result for the remediated output

Notes

• JSON object order is for readability in this proposal. Implementations should not rely on key order.
• The report should avoid raw PHI, credentials, secrets, or full sensitive snippets unless explicitly approved for a controlled environment.
• If remediation fails or is not eligible, remediation.applied should be false and no remediated output should be returned.
• For multimodal requests, input.normalized_artifacts can include transcripts, image derivatives, OCR output, or sampled video frames when policy allows them.