CONFIDENCE GATING
Below its threshold, the agent does not act. It escalates with what it saw and why it is unsure. Silence is a valid output.
AI AGENTS
Agents that
touch metal.
A Serumon agent is not a chatbot with a plant integration. It is a constrained control loop: a learned world model of your formula and vessel, a policy clamped by validated envelopes, a human checkpoint, and an immutable record of every decision it made and why.
4AUTONOMY RUNGS
100%CITED OUTPUTS
BOUNDEDEVERY ACTION
REPLAYABLEEVERY DECISION
{
"agent": "emulson",
"vessel": "V-07", "batch": "B-24118",
"observation": { "visc_pred": 43120, "d50": 2.44 },
"prediction": { "eob_visc": 44900, "risk": 0.31 },
"proposal": { "shear_rpm": -40, "hold_s": 90 },
"envelope": { "shear_min": 2900, "shear_max": 3400 },
"confidence": 0.91,
"citations": ["SOP-114 §4.2", "batch B-23907"],
"approved_by": "a.rossi (process eng)",
"model": "emulson-wm-2026.07.03",
"result": "in spec"
}WORLD MODELMPCRL POLICYPINN SURROGATERAG + CITATIONSCONFIDENCE GATINGHUMAN CHECKPOINTIMMUTABLE AUDIT
ANATOMY
What is inside one agent
PERCEPTION
Multimodal encoders over inline sensors, spectra, images and batch context, running quantised on the edge GPU.
WORLD MODEL
A learned model of how this formula behaves in this vessel: what happens to viscosity, D50 and shade if you do X now.
POLICY
MPC or RL policy proposing the next control move, evaluated against the predicted end-of-batch state.
ENVELOPE
A hard constraint layer. A proposal outside the envelope is not actioned, it is escalated. No exceptions, no override in code.
CHECKPOINT
The human gate appropriate to the current autonomy rung: observe, approve, or escalate-only.
RECORD
A signed decision record with observation, prediction, proposal, confidence, citations, approver, model version and outcome.
GRADUATED AUTONOMY
Four rungs, earned per formula
-
01
SHADOW
The agent observes the vessel or line and predicts. It writes nothing. You compare its calls against your chemists and your lab for the whole baseline period.
-
02
RECOMMEND
The agent proposes the dose, shear, shade or fill move with its reasoning and evidence. An engineer accepts or rejects; every rejection trains the model.
-
03
APPROVE-TO-ACT
The agent stages a control move inside an approved envelope. A named operator releases it. Every action is signed into the immutable audit log.
-
04
BOUNDED AUTONOMY
Inside validated envelopes and on qualified formulas the agent acts on its own, escalating anything outside the envelope to a human within one control cycle.
SAFETY DESIGN
How we keep an agent from doing something stupid
ENVELOPE CLAMPING
Actuation limits are enforced outside the model, in the runtime, so a model error cannot become an out-of-range write.
RATE LIMITING
Not just how far a parameter can move, but how fast and how often. Oscillation is a failure mode we design against.
SAFE-STATE FALLBACK
Watchdog failure, sensor loss or node failure reverts control to your existing DCS recipe within one cycle.
DUAL APPROVAL
Envelope changes, model promotions and autonomy-rung changes require two named approvals, recorded.
REPLAY + POST-MORTEM
Any batch can be replayed against the exact model and envelope, which makes root cause a query rather than an argument.
MODEL INVENTORY
Which model does what, and how real it is
| MODEL | JOB | STATUS |
|---|---|---|
| Process world model | Predict emulsion state and end-of-batch spec | Real for design-partner formulas; fleet generalisation aspirational |
| Rheology / fault predictor | Viscosity, D50 and emulsion-fault risk | Real |
| Spectral ΔE regression | Shade against brand standard | Real |
| Defect vision | Multi-format cosmetic defect classification | Real |
| Fill fusion + trim | Giveaway optimisation under compliance | Real |
| Robot manipulation policy | Assembly, capping, packout | Real in simulation; on-line per-cell validation pending |
| Micro / stability risk | Predictive batch safety | Real where data exists; lead-time claims aspirational |
| PINN surrogates | Interactive emulsification simulation | Real; full-fidelity twin aspirational |
| Grounded reasoning | Deviation write-ups, SOP Q&A | Real, citation-enforced |
GROUNDING
No answer without a source
Language models in this platform are never the decision-maker on a control action. They read SOPs, batch records and specs, and they draft explanations and investigations for a human to verify.
Every generated statement carries citations to the artefacts that support it. If retrieval finds nothing sufficient, the correct output is a refusal, not a plausible paragraph. In a GMP environment, a confident hallucination is a deviation.
- Per-tenant vector stores with permission-aware filtering.
- Citations enforced at generation time, not bolted on.
- Refusal when evidence is insufficient.
- Model-agnostic router across frontier and fine-tuned open models.
- Self-hosted option via NIM for sites that cannot send data out.
GROUNDING POLICY
- UNCITED CLAIMS
- BLOCKED
- INSUFFICIENT EVIDENCE
- REFUSE
- CONTROL ACTIONS FROM LLM
- NEVER
- SOP RETRIEVAL
- PERMISSION-AWARE
- ON-PREM OPTION
- AVAILABLE
- AUDIT OF PROMPTS
- RETAINED
TRAINING LOOP
How the agents get better
STEP 01
CAPTURE
Every batch produces telemetry, spectra, images and outcomes at full resolution on the edge.
STEP 02
LABEL
Engineer approvals, rejections and corrections become labels with reason codes attached.
STEP 03
TRAIN
Models are retrained on DGX-class capacity, evaluated against held-out batches per formula family.
STEP 04
PROMOTE
Signed rollout, canary on one vessel, measured, then fleet promotion or rollback.
AGENT SLAs
The numbers the runtime holds
<150msP95 DECISION
1 CYCLEESCALATION LATENCY
100%DECISIONS RECORDED
2APPROVALS FOR CHANGE
ENGINEERING TARGETS FOR DESIGN-PARTNER DEPLOYMENTS.
Trust before autonomy. Grounding, audit, human-in-the-loop, graduated control — in that order, every time.
AGENT FAQ
How the autonomy actually behaves
In shadow and recommend modes, the chemist wins and the disagreement is recorded as a training signal. Over a pilot, the pattern of disagreements is the single most useful artefact produced — it tells both sides where the model is weak and where the craft is undocumented.
Yes. Autonomy is a per-formula, per-vessel switch with an immediate effect and a recorded actor. Reverting drops you to recommend mode; the agent keeps predicting and keeps learning.
Not with your formulas or shade standards. Cross-tenant learning is limited to architectural improvements and explicitly consented, anonymised benchmark products. Formula IP protection is a contractual commitment, not a setting.
Every model version is immutable and tied to a validation record. Changes flow through your change control, roll out as canaries on a single vessel, and can be reverted with one action. Any batch can be replayed against the version that ran it.
GO DEEPER
Interrogate the architecture
Bring your process engineers and your quality lead. The good questions are the hostile ones, and we would rather answer them before a contract than after.