Putting the sovereign model to work on the energy island

• A proprietary SAVRN model, trained and served by SAVRN on SAVRN hardware. No external model provider is in the path.
• One model path: every SAVRN module reaches the model through the SAVRN AI Gateway. The gateway enforces policy, logs every call, and refuses external providers. Gateway policy is platform policy.
• Tool calling runs through the gateway: the model calls plant tools, and every number it states comes back from a tool result.
• Training program: the model is trained on the plant corpus in rounds, qualified against a golden set of plant tasks before it goes live, and cut over with a manifest and a rollback.
• Separate lanes: production agents and operators use the SAVRN model only. Evaluation against outside models happens in a separate lane that never touches plant tools.

Nine parts. Click any part to open it.

The gas line on the property feeds the combined-cycle gas block through a custody-transfer meter station and a conditioning skid.

The generators are the combined-cycle gas block (gas turbine plus steam turbine), the 100 MW steam turbine (3.12), and the two biomass turbine-generators (3.10). Each comes with an OEM controller that publishes the tags below.

The HRSG on the gas block supplies extraction steam to the steam header and exhaust heat to the absorption chillers.

• Utility: tariff, interconnection agreement terms, curtailment or export rules, outage notices.
• Gas supplier: nominations, daily balancing, price index, operational flow orders.
• Permits: state license conditions, air district permit limits, Title V obligations, GHG reporting calendar.
• O&M: OEM manuals, procedures, P&IDs, one-lines, alarm rationalization documents, CMMS history.

These documents are the first training and retrieval corpus for the model. They exist before the plant runs, so the model learns the plant first. See section 07.

Two wood-waste steam plants, 18 MW and 21 MW. They are the §45Y facility and the CO2 source for capture, so their measurements carry tax-credit weight as well as operating weight.

Refractory brick charged by resistive heaters on off-peak electricity, discharging continuous steam. When it charges and what it pays is an optimization problem. Its thermal state of charge decides whether the reboiler and the turbine have steam.

The capture unit turns the biomass stack into the §45Q and removal-attribute source. Its metering is regulatory. The reporting rules for geologic sequestration (40 CFR 98 Subpart RR for injection, Subpart PP for supply) and the Class VI well program define what must be measured and how.

# Tool definitions (OpenAI-style function calling, as the SAVRN gateway already uses)
[
  {"type":"function","function":{"name":"query_tags",
    "description":"Query the plant historian. Returns aggregated time series for named tags.",
    "parameters":{"type":"object","properties":{
      "tags":{"type":"array","items":{"type":"string"},"maxItems":24},
      "from":{"type":"string","description":"ISO-8601 start"},
      "to":{"type":"string","description":"ISO-8601 end"},
      "agg":{"type":"string","enum":["avg","min","max","last"]},
      "bucket":{"type":"string","enum":["1m","5m","15m","1h","1d"]}},
      "required":["tags","from","to"]}}},
  {"type":"function","function":{"name":"get_active_alarms",
    "description":"List active alarms with priority, first-out flag, and asset.",
    "parameters":{"type":"object","properties":{"min_priority":{"type":"integer"}}}}},
  {"type":"function","function":{"name":"get_unit_status",
    "description":"Rolled-up status for generation units, BESS, chillers, BTES arrays, intertie.",
    "parameters":{"type":"object","properties":{}}}},
  {"type":"function","function":{"name":"search_docs",
    "description":"Retrieve passages from O&M manuals, permits, procedures, P&IDs, interconnection agreement.",
    "parameters":{"type":"object","properties":{"q":{"type":"string"},"k":{"type":"integer","default":6},
      "collection":{"type":"string","enum":["manuals","permits","procedures","drawings","agreements"]}},
      "required":["q"]}}},
  {"type":"function","function":{"name":"get_forecast",
    "description":"Load, thermal, or weather forecast from the site forecasters.",
    "parameters":{"type":"object","properties":{"kind":{"type":"string","enum":["load","thermal","weather"]},
      "horizon":{"type":"string","enum":["1h","24h","7d"]}},"required":["kind","horizon"]}}},
  {"type":"function","function":{"name":"get_dispatch_plan",
    "description":"Latest optimizer unit-commitment and heat-routing plan with constraints.",
    "parameters":{"type":"object","properties":{"date":{"type":"string"}}}}},
  {"type":"function","function":{"name":"file_advisory",
    "description":"Store a report or recommendation with its evidence trace.",
    "parameters":{"type":"object","properties":{"kind":{"type":"string"},"title":{"type":"string"},
      "body_md":{"type":"string"},"evidence":{"type":"array","items":{"type":"string"}}},
      "required":["kind","title","body_md","evidence"]}}},
  {"type":"function","function":{"name":"request_setpoint",   # step 4 only; disabled until then
    "description":"Propose a bounded change to an allow-listed setpoint. Requires human approval.",
    "parameters":{"type":"object","properties":{"setpoint_id":{"type":"string"},
      "value":{"type":"number"},"justification":{"type":"string"},"ttl_minutes":{"type":"integer"}},
      "required":["setpoint_id","value","justification"]}}}
]

# Same path every SAVRN module uses: the SAVRN AI Gateway
POST {gateway}/v1/chat
{
  "model": "savrn-plant-ops",
  "messages": [
    {"role":"system","content":"You are the SAVRN energy-island operations assistant. Every number you state must come from a tool result. You do not control equipment."},
    {"role":"user","content":"Write the 06:00 shift handover for the gas island and thermal plant."}
  ],
  "tools": [ ...the list above... ],
  "tool_choice": "auto"
}
# The model calls get_unit_status, query_tags (fuel, MW, heat balance, BTES), get_active_alarms,
# then file_advisory with kind="shift_report". Nothing else happens.

Scripts that poll OPC UA from the model-facing layer, paste raw numbers into prompts, or post to an ad hoc local model server are not used. These three rules apply.

Minimal example

# Tier 3: tool service over the historian (FastAPI), and a call through the gateway with tools.
# Edge (Tier 2) already moved OPC UA / Modbus data into TimescaleDB via Sparkplug B.
from fastapi import FastAPI, Query
import psycopg, requests, json

app = FastAPI()

@app.get("/v1/tags/query")
def query_tags(tags: list[str] = Query(..., max_length=24), frm: str = "", to: str = "",
               agg: str = "avg", bucket: str = "5m"):
    sql = f"""SELECT time_bucket(%s, ts) AS t, tag, {agg}(value) AS v
              FROM plant_ts WHERE tag = ANY(%s) AND ts BETWEEN %s AND %s
              GROUP BY t, tag ORDER BY t LIMIT 5000"""
    with psycopg.connect("dbname=plant") as cx:
        rows = cx.execute(sql, (bucket, tags, frm, to)).fetchall()
    return {"bucket": bucket, "agg": agg, "rows": rows}

# Caller (a SAVRN agent or a scheduled job): gateway + tools, never raw numbers.
def shift_report():
    r = requests.post(GATEWAY + "/v1/chat", json={
        "model": "savrn-plant-ops",
        "purpose": "plant_ops",
        "messages": [
            {"role": "system", "content": "You are the SAVRN energy-island operations assistant. "
                                          "State only numbers returned by tools. You do not control equipment."},
            {"role": "user", "content": "Write the 06:00 shift handover for the gas island and thermal plant."}],
        "tools": PLANT_TOOLS,          # the list in this section
        "tool_choice": "auto"})
    return r.json()   # the gateway runs the tool loop against the Tier 3 service and returns the advisory

Tool results are structured objects keyed by namespace path, not flat tag dumps. The model is trained on this shape, so it reads meaning and status, not just a number.

{
  "path": "site/biomass_capture/amine_unit/reboiler_duty_mw",
  "asset": "Amine capture unit",
  "measure": "Reboiler thermal duty",
  "value": 41.2, "unit": "MW_th",
  "status": "nominal",          # nominal | watch | alarm, from the alarm server
  "quality": "good",            # good | uncertain | bad, from the edge
  "ts": "2026-08-21T13:05:00Z",
  "range": [30, 50], "source": "capture DCS via OPC UA"
}

The model populates a draft. A deterministic validator checks it against the engineered envelope. An operator approves it in the HMI. Only then does anything move.

# draft from the model (via request_setpoint)
{"setpoint_id":"rondo.charge_power_mw","value":120,"ttl_minutes":60,
 "justification":"Off-peak price window 02:00-06:00; thermal SoC 38 percent; forecast reboiler + turbine demand 190 MW_th."}

# envelope held by the setpoint service, written by the engineer of record, changed only by change control
{"setpoint_id":"rondo.charge_power_mw","min":0,"max":170,"ramp_limit_mw_per_min":10,
 "allowed_modes":["grid_connected"],"requires_approval":true,"approver_role":"shift_supervisor"}

# validator result: pass or a named reason; every request and outcome is logged to the historian
{"request_id":"...","valid":true,"checks":["range","ramp","mode","ttl"],"approval":"pending"}

• A purpose value, purpose="plant_ops", maps to the plant model, local only, tools allowed, and the plant tool set.
• request_setpoint is not in the tool list until step 4, and the endpoint returns 403 until the setpoint service is commissioned.
• Every tool call and result is logged with the conversation id, so an advisory can be replayed and audited.

Goal: be able to draw the plant's zones and say which protocol crosses which boundary.

• NIST SP 800-82 Rev. 3, Guide to OT Security (free): csrc.nist.gov/pubs/sp/800/82/r3/final. Read chapters on architecture and the Purdue reference.
• ISA/IEC 62443 overview from ISA: isa.org (62443 series).

Lab 1: redraw the ring diagram in section 02 for the energy island with every subsystem in section 03 placed in a zone.

Goal: know what a transmitter, a flow computer, a CT/PT, and a vibration probe actually put on the wire.

• Kuphaldt, Lessons in Industrial Instrumentation (free, creative commons): ibiblio.org/kuphaldt/socratic/sinst/. Chapters on pressure, temperature, flow, level, and 4 to 20 mA / HART.
• AGA Report No. 3, 8, 9 summaries from the flow-computer vendor of your choice (Emerson, ABB, Honeywell all publish primers).

Lab 2: build the tag dictionary for subsystems 3.1 and 3.2: name, unit, range, rate, source protocol. This file becomes training data.

Goal: read a register, browse an OPC UA server, publish a Sparkplug metric.

• Inductive University (free video courses on Ignition, SCADA, MQTT): inductiveuniversity.com.
• Eclipse Sparkplug specification: sparkplug.eclipse.org.
• OPC Foundation, OPC UA overview: opcfoundation.org.
• Python: pymodbus, asyncua, paho-mqtt.

Lab 3: on a workstation run Mosquitto (broker) and a Python script that simulates the gas turbine (MW, Hz, exhaust temperatures, fuel flow) as a Modbus TCP server, a second script that polls it and publishes Sparkplug B, and a third that subscribes and prints. You now have the edge layer of section 04.

Goal: store tags in TimescaleDB, query by time bucket, attach the asset model.

• TimescaleDB docs (hypertables, continuous aggregates): docs.timescale.com.
• Grafana for trends: grafana.com/docs.

Lab 4: subscribe the Lab 3 stream into TimescaleDB, build a 1-minute continuous aggregate, and plot it in Grafana. Write the /v1/tags/query endpoint in FastAPI over it.

Goal: understand governor and AVR roles, droop vs isochronous, grid-forming vs grid-following, and what the relay device numbers mean.

• Blackburn and Domin, Protective Relaying: Principles and Applications (textbook).
• IEEE 1547-2018 and IEEE 2800-2022 summaries (IEEE SA): standards.ieee.org.
• SEL application guides and the SAVRN power management and controls drawing for the PMC and BESS pattern.
• Woodward governor fundamentals (vendor primers on speed and load control).

Lab 5: extend the simulator with a grid-forming BESS (SoC, P/Q, mode) and an intertie relay state. Simulate a GPU step load and watch BESS take it while the turbines ramp. Write up which layer acted and how fast, using the control timing table in section 02.

Goal: do the heat balance by hand: fuel in, electricity out, reboiler duty, Rondo steam, turbine extraction, chiller tons, BTES charge.

• The Rondo heat battery datasheet for charge, storage, and discharge behavior.
• ASHRAE Handbook, HVAC Systems and Equipment, chapter on absorption equipment.
• The SAVRN site development document (heat-balance table) and the BTES section drawing.
• IEA ECES / borehole thermal storage primers; thermal response test method references.

Lab 6: add Rondo (charging MW, brick temperatures, steam out), the steam header, the amine reboiler duty, and the thermal plant (exhaust heat, absorption chiller COP, CDU loop, BTES array temperatures with a slow seasonal drift) to the simulator. Expose /v1/units/status for it.

Goal: know what a CEMS reports, what the averaging periods are, and what the permit limit looks like as data.

• EPA, 40 CFR Part 60 Appendix B (performance specifications) and Appendix F (QA): ecfr.gov (40 CFR 60).
• The local air district rules and the state power plant licensing pages.

Lab 7: simulate CEMS 1-minute NOx and O2, compute the 15 percent O2 corrected value and the permit averaging period, and expose /v1/compliance/cems. Have the model draft an excess-emission report from a synthetic exceedance.

Goal: build a simple load forecaster, a statistical anomaly detector, and a small dispatch optimizer, each behind an endpoint.

• Forecasting: gradient-boosted trees on scheduler queue + weather (scikit-learn or LightGBM).
• Anomaly detection: rolling z-score and isolation forest as a start.
• Optimization: a small MILP with PuLP or OR-Tools for unit commitment and heat routing.

Lab 8: ship /v1/forecast/load and /v1/dispatch/plan from these. Keep them deterministic and versioned.

Goal: run jobs 1 to 5 end to end on the simulator through the SAVRN AI Gateway with the SAVRN model.

• SAVRN model serving and gateway documentation, the training pipeline, and the qualification scripts.

Lab 9: register the tool set in the gateway under purpose=plant_ops, index the corpus in section 07, run the golden set, train round 1, compare against the untrained base model on the same tool path. This is the step 2 done-when check.

Goal: know what the model path costs in time and memory, and where that matters. Alerting is the alarm system's job and is not on the model's clock. The model's latency matters for the operator copilot (a few seconds is fine) and for scheduled jobs (minutes are fine).

• Inference server options: context size, KV cache quantization, parallel slots, prompt caching.
• Quantization trade-offs: measure before changing.
• Prompt caching for the long system prompt and tool definitions, so a tool-loop turn costs only the new tokens.

Lab 10: measure time-to-first-token and tokens per second for a shift-report call with the full tool list, with and without prompt caching. Record the budget. Size the plant inference node from it.