Syzygy Rosetta - Origin Codebase

API-first AI governance middleware MVP for evaluating submitted content through deterministic rules, safety tags, and risk scoring.

What Is Syzygy Rosetta?

Syzygy Rosetta is a FastAPI-based governance decision service. The current API accepts a required user/customer input, an optional model output, and optional context. It evaluates the interaction through safety tagging, deterministic policy rules, and risk scoring, then returns a structured decision:

• allow
• rewrite
• escalate

user input + optional model output -> POST /evaluate -> allow / rewrite / escalate
                                                  |
                                                  +-> safety_layer.py
                                                  +-> config/policy_rules.json
                                                  +-> core/reflex.py
                                                  +-> logs/evaluations.json

API Response

Every successful POST /evaluate response returns these 8 fields:

{
  "decision": "allow | rewrite | escalate",
  "risk_score": 0.12,
  "confidence": 0.91,
  "violations": [],
  "rewrite": null,
  "reasoning": "Interaction evaluated as low risk. Continue with normal processing.",
  "field_notes": [],
  "timestamp": "2026-03-21T14:32:00Z"
}

Decision Thresholds

Risk Score	Decision	Behavior
< 0.4	allow	Interaction passes. violations is empty.
0.4 - 0.7	rewrite	Input or output should be clarified or rewritten. rewrite is populated.
>= 0.7	escalate	Interaction is routed to human review. rewrite is null.

File Structure

syzygy-rosetta-originbase/
|
|-- README.md
|-- syzygy-rosetta-originbase.md
|-- REPO_MISMATCH_AUDIT.md
|
`-- syzygy-rosetta/
    |-- app.py                     FastAPI entry point
    |-- run_api.py                 local development launcher
    |-- safety_layer.py            pre-classification tags and sensitive-topic detection
    |-- requirements.txt           Python dependencies
    |-- Dockerfile                 container entrypoint for app:app
    |
    |-- config/
    |   `-- policy_rules.json      deterministic industry rules
    |
    |-- core/
    |   |-- reflex.py              governance decision engine
    |   |-- risk_scoring.py        weighted feature scoring utilities
    |   |-- constants.py           invariant/config constants
    |   |-- invariants.json        invariant metadata
    |   `-- resonators_mock.py     legacy/simple reflex mock
    |
    |-- docs/
    |   `-- demo_checklist.md
    |
    |-- example/
    |   `-- basic_usage.py
    |
    |-- logs/
    |   `-- evaluations.json       runtime audit log
    |
    `-- tests/
        |-- test_evaluate.py
        `-- test_healthz.py

Evaluation Pipeline

Every POST /evaluate call currently follows this path:

1. FastAPI validates required input, optional output, and optional context.
2. evaluate_prompt() runs a breath pause and mirror step.
3. safety_layer.tag_input() labels authority, manipulation, dependency, and escalation patterns on input and output.
4. detect_sensitive_topic() checks self-harm, violence, and sexual-content patterns on input and output.
5. _apply_policy_rules() checks industry-specific rules from config/policy_rules.json on input and output.
6. The active scorer computes risk and confidence for the input-output pair.
7. Risk floors and multipliers are applied.
8. The API returns the 8-field response and appends an entry to logs/evaluations.json.

Industry Context

Pass industry in the request context to activate sector-specific policy rules:

Industry	Policy Rules
finance	Flags coercive financial instructions, guaranteed-return claims, compliance bypass, and market misconduct.
healthcare	Flags unsafe medication directives, system override attempts, and unauthorized access.
general	Flags jailbreak attempts, system prompt injection, unsafe security bypass requests, and harmful instructions.

Production environment multiplier: x1.10. Multiple violations multiplier: x1.15.

Quick Start

Requirements

• Python 3.11+
• Docker Desktop, only if running the container

Run Locally

Run these commands from syzygy-rosetta-originbase/syzygy-rosetta:

python -m pip install -r requirements.txt
python run_api.py

The local server starts at:

http://127.0.0.1:8000

Run With Docker

The Dockerfile currently lives inside syzygy-rosetta/, so build from that directory:

cd syzygy-rosetta
docker build -t rosetta .
docker run -p 8000:8000 rosetta

Test The Endpoint

curl -X POST http://localhost:8000/evaluate \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Summarize the key risks in this portfolio.",
    "output": "The portfolio appears diversified, but review concentration and liquidity risks.",
    "context": {
      "environment": "staging",
      "industry": "finance"
    }
  }'

Example response shape:

{
  "decision": "allow",
  "risk_score": 0.14,
  "confidence": 0.5,
  "violations": [],
  "rewrite": null,
  "reasoning": "Interaction evaluated as low risk. Continue with normal processing.",
  "field_notes": [
    "FIELD_NOTE [2026-04-25T16:41:48Z]: mirror invoked",
    "INTERNAL_NOTE [2026-04-25T16:41:48Z]"
  ],
  "timestamp": "2026-04-25T16:41:48Z"
}

Explore Via Swagger UI

http://localhost:8000/docs

Evaluation Log

Every POST /evaluate call appends one entry to syzygy-rosetta/logs/evaluations.json:

{
  "timestamp": "2026-03-21T14:32:00Z",
  "input": "the original input string",
  "output": "the model output string or null",
  "decision": "allow | rewrite | escalate",
  "risk_score": 0.85,
  "confidence": 0.91,
  "violations": ["violation_label"],
  "rewrite": "rewritten string or null",
  "reasoning": "decision explanation",
  "field_notes": [],
  "context": {
    "user_id": "string or null",
    "environment": "production | staging",
    "industry": "finance | healthcare | general"
  }
}

Run Tests

Run from syzygy-rosetta-originbase/syzygy-rosetta:

python -m pytest tests -q

At the time of the audit, this produced:

54 passed

Repository Role

This is the origin codebase for Syzygy Rosetta. It contains the foundational MVP implementation of the governance engine. Active development and refactoring are ongoing.

For known mismatches and cleanup work, see REPO_MISMATCH_AUDIT.md.

License

Licensed under AGPL-3.0-or-later.

Derived from the Syzygy Rosetta v1.0 protocol by Sarasha Elion (Trivian Institute).