AI Products

Releases & Comparison

Every time you ship a version of an AI product, BRICQS creates an immutable release snapshot. The release stores everything about the product at that point in time — model config, capabilities, safety score, and deployment strategy — enabling exact comparison between any two releases.

What a release contains

Version string
Semantic version (e.g. v2.1.0), stored immutably on creation
Release status
draft → evaluation → review → preview → approved → production → archived
Model config
Model ID, temperature, system prompt, and tool definitions — exact values, not references
Capabilities
List of active capabilities at release time with dependencies resolved
Safety score
Numeric score from the last evaluation run before release
Snapshot hash
SHA-256 of the full release payload — proves the snapshot is unchanged
Deployment strategy
Recommended environment, min/max replicas, and rollback policy for this release
Traffic weights
Traffic split for canary/blue-green deployments (release → production ratio)

Release comparison (diff)

BRICQS automatically computes a structured diff between consecutive releases and exposes it via API. The diff is computed when a release transitions to approved status. You can also trigger a comparison between any two arbitrary releases.

Diff categories
behavior_change

System prompt changes, temperature changes, tool definition changes — anything that would alter the model's output behavior.

data_change

Changes to the knowledge base, retrieval configuration, or embedded documents used for RAG.

infrastructure_change

GPU vs CPU runtime switch, replica count changes, environment variable additions or removals.

schema_change

Input/output schema changes, capability additions or removals, API contract changes.

safety_change

Safety score delta between releases. Positive delta means the new release scored higher on evaluation.

API

GET/releases/{release_id}/diff

Returns the structured diff between this release and its predecessor. Secret values are never included in the diff — only the presence/absence of a secret key is reported.

{
  "release_id": "rel_abc",
  "compared_to": "rel_xyz",
  "behavior_changes": [
    {
      "field": "system_prompt",
      "before": "You are a contract analyst...",
      "after": "You are a precise contract analyst..."
    }
  ],
  "capabilities_added": ["document_comparison"],
  "capabilities_removed": [],
  "model_before": "gpt-4o",
  "model_after": "gpt-4o",
  "safety_score_before": 0.82,
  "safety_score_after": 0.91,
  "deployment_strategy": "canary_10_percent",
  "breaking_changes": false
}
POST/releases/{id}/status

Advance a release through its status lifecycle. Only valid transitions are accepted — the API returns 400 if the transition is not valid from the current status.

Viewing diffs in the console

Navigate to Console → AI Projects → [project] → Releases → [release] → Compare. The Compare tab shows a GitHub PR-style diff with before/after for each changed field, colored capability badges, safety score delta, and the recommended deployment strategy for the new release.