SystemOne API

Production API for the Security-One deterministic structured-classification model. Its public wire contract is compatible with OpenJev's POST /v1/systemone: one shared state, any mix of noul, choice, and score questions, and answers returned under the caller's question IDs.

The production deployment serves the verified AutoJev Security V2.3 release as security-one:

• frozen Qwen3.8-27B BF16 checkpoint with an exact 255-way AutoJev readout export;
• zero class flips and zero logit drift in the 64-example export gate;
• fixed probability calibration temperature 0.14527332485151376;
• SGLang selected-token log probabilities, with one-token deterministic answer labels;
• shared-state prefix reuse for parallel questions;
• a 65,536-token branch context and no silent truncation.

Modal endpoint after the paused service is redeployed: https://superagentai--systemone-api-web.modal.run

API reference after redeployment: https://superagentai--systemone-api-web.modal.run/

API

POST /v1/systemone
Authorization: Bearer <key>
Content-Type: application/json

curl "$SYSTEMONE_URL/v1/systemone" \
  -H "Authorization: Bearer $SYSTEMONE_API_KEY" \
  -H "Content-Type: application/json" \
  --data @examples/request.json

state may be a string, object, or array. Every question sees the same state and is evaluated in parallel.

• noul is binary and returns the probability of true as noul.
• choice returns the highest-probability option, the full distribution, and normalized-entropy confidence.
• score returns the expected zero-based level, a legend, the full distribution, and confidence.

The aliases security-one-latest, systemone, systemone-latest, openjev, and jev-latest all select the same loaded release. If model is omitted, it defaults to security-one.

Additional endpoints:

• GET /models — public OpenRouter provider document (schema 2.4)
• GET /v1/models
• GET /v1/limits
• GET /health for inference readiness
• GET /health/live for process liveness
• GET /docs, /openapi.json, and / for API reference

Local API development

The API process expects a running SGLang server with the selected-token logprob endpoint enabled.

uv sync --extra dev
cp .env.example .env
uv run systemone serve

Set SYSTEMONE_MODEL_PATH to the local tokenizer snapshot and SYSTEMONE_BACKEND_URL to SGLang. Local auth is disabled only when both static-key settings and the customer lookup URL are empty. Static-only local development remains supported. Modal forces production mode and refuses to start without customer-key validation configuration.

Run the checks with:

uv run --extra dev pytest
uv run --extra dev ruff check .

Production on Modal

The deployment is pinned to the verified autojev-security-v2-3-sglang-export-20260925a artifacts in the private openjev-rlcd-artifacts-v1 volume. It starts one SGLang server and the API in the same B200 container, so the backend is never exposed publicly.

The prepared deployment accepts Superagent organization sk_live_ keys through an authoritative, indexed SHA-256 lookup on every request. It preserves SYSTEMONE_API_KEYS and SYSTEMONE_API_KEYS_SECONDARY for internal/testing use. Authentication happens directly in Modal; there is no Next.js authentication hop or customer-key synchronization.

Before deploying, follow authentication setup and rollout: provision a dedicated read-only PostgreSQL role when possible, and a server-side Modal Secret systemone-customer-auth. Current production uses the documented server-only Supabase fallback because Supabase-managed PUBLIC permissions prevent a narrow direct-login role. Its service-role credential has broad project access; see the runbook's residual-risk section. Keep systemone-production and systemone-testing-key secrets unchanged. Do not deploy or change live secrets without approval. After approval and configuration, deploy:

uv run --extra modal modal deploy deployment/modal_app.py

The initial early-access deployment keeps exactly one warm B200 container and caps scaling at one to bound spend. Raise max_containers only after load testing; lower min_containers only when cold starts are acceptable.

Verify the live endpoint, including all three answer types:

SYSTEMONE_API_KEY="<customer-key>" uv run systemone smoke https://YOUR-ENDPOINT.modal.run

Operational behavior

• Static keys are compared in constant time; organization keys use a reused async client/pool and uncached hash lookup. Keys, hashes, credentials, and organization identities are not logged.
• State, questions, and prompts are never written to application logs.
• Request bodies, question counts, option counts, per-branch context, total tokens, concurrency, and deadlines are bounded before or during admission.
• Overload, timeout, and backend failures return 503 with Retry-After: 1.
• Invalid/deleted keys return 401; authentication outages fail closed with 503 before reading the body or running inference. Authentication is measured separately from inference.
• Each response includes x-systemone-request-id, x-systemone-model, cache diagnostics, and Server-Timing without exposing customer data.
• usage.output_tokens reports the actual internal warm-up plus one-token decisions. Billing can still price output at zero; the usage record is not falsified.

Release checklist

1. Run the unit and contract tests.
2. Build the Modal image; its fail-closed source patches must apply exactly once.
3. Smoke-test all three question types against the deployed URL.
4. Run the frozen quality canary before moving customer traffic.
5. Verify created/deleted organization keys, static fallback, public routes, and lookup outages; benchmark p50/p95 lookup and end-to-end latency.
6. Put the endpoint behind the production domain/rate limiter before broader access.
7. Release the landing model API docs after hosted organization authentication is verified.

Research results

• SystemOne v4 multi-objective continuation
• SystemOne v3 general-reasoning continuation
• SystemOne v2 general-classifier experiment

The v4 package is the latest research evaluation, not the currently deployed production release. It contains aggregate results and reproducibility metadata only; training rows, teacher outputs, model weights, credentials, and customer data are not published.