DOCUMENTATION

Read before
you wire.

The operating manual for a system that writes setpoints into your plant: architecture, deployment, connectors, envelopes, autonomy, audit, model lifecycle and runbooks. Full documentation is provisioned per deployment; this is the public map.

8SECTIONS
PER-DEPLOYRUNBOOKS
VERSIONEDWITH THE PRODUCT
EXPORTABLEFOR VALIDATION
AUTONOMY LADDER4 RUNGS
SHADOW RECOMMEND APPROVE BOUNDED READ ONLYPROPOSES SIGN-OFFIN LIMITS EVIDENCE REQUIRED →

No rung is skipped. Any rung can be revoked in one action.

ARCHITECTUREDEPLOYMENTCONNECTORSENVELOPESAUTONOMYAUDITMODEL LIFECYCLERUNBOOKS

GETTING STARTED

The first four documents

DOC 01

ARCHITECTURE

How the edge runtime, model stack, data platform and console fit together, with data-flow diagrams for security review.

DOC 02

SITE SURVEY

What we need to see: vessels, homogenisers, spectros, fillers, robots, MES, historian, network segmentation and identity.

DOC 03

INSTALLATION

Edge node placement, power and network requirements, K3s bootstrap, certificate provisioning and connectivity validation.

DOC 04

FIRST CONNECTOR

Tag discovery, canonical schema mapping, unit declaration, range validation and replay testing before go-live.

ENVELOPES

The most important document you will write

An envelope defines what an agent is allowed to do to a parameter: minimum, maximum, maximum rate of change and maximum frequency, per formula and per vessel.

It is a controlled document with dual approval and full version history, because in practice it is the contract between your process owners and the autonomy.

  • One envelope per parameter, per formula, per vessel.
  • Min, max, rate limit and frequency limit are all mandatory.
  • Dual approval required to create or change.
  • Pinned to every action in the audit trail.
  • Diffable and reversible like any other controlled artefact.
envelope.yaml
formula: SPF50-rev14
vessel: V-07
parameters:
  homogeniser_rpm:
    min: 2900
    max: 3400
    max_delta_per_move: 60
    min_seconds_between_moves: 45
  jacket_setpoint_c:
    min: 74.0
    max: 80.0
    max_delta_per_move: 0.5
approvals:
  - process_engineering: a.rossi
  - quality: m.chen

AUTONOMY

Promoting a formula up the ladder

  1. 01

    EVIDENCE

    Accumulate shadow and recommend performance for the formula: prediction error, agreement rate with chemists, and near-miss analysis.

  2. 02

    REVIEW

    Process and quality review the evidence pack and the proposed envelope together, in one meeting, with the model version fixed.

  3. 03

    PROMOTE

    The autonomy rung is raised for that formula and vessel only, with dual approval and an immediate audit entry.

  4. 04

    MONITOR

    Elevated monitoring for a defined number of batches, with automatic demotion on threshold breach.

MODEL LIFECYCLE

From training to a vessel

TRAIN

DGX-class training over telemetry, spectra, images and outcomes, evaluated per formula family on held-out batches.

SIGN

Artefacts are signed and immutable. Unsigned models cannot load on an edge node under any configuration.

CANARY

Deployed to one vessel or line class, monitored against defined acceptance criteria for a defined period.

PROMOTE / REVERT

Fleet promotion or single-action rollback. Every decision record keeps its model version forever.

RUNBOOKS

What to do when something is wrong

SENSOR LOSS

Detection, degraded-mode behaviour, which predictions are suppressed, and how control reverts.

NODE FAILURE

Safe-state transition, DCS handback, spare-node procedure and how to resume without replaying stale commands.

MODEL REGRESSION

Detection thresholds, immediate demotion, rollback procedure and the post-incident report template.

ENVELOPE BREACH

What the runtime does, who is notified, how the attempt is recorded and the investigation checklist.

AUDIT REQUEST

How to export a full batch replay for an inspector, including model and envelope versions.

CONNECTIVITY LOSS

Offline behaviour, local buffering limits, reconciliation on reconnect and what is never replayed.

REQUIREMENTS

What a site needs before install

What a site needs before install
REQUIREMENTMINIMUMNOTE
Edge nodeJetson Orin / IGX class per vessel clusterSized during survey by sensor and camera load
NetworkSegmented OT VLAN, outbound egress policyNo inbound internet access required
Equipment accessOPC-UA or Modbus with documented tagsRead-only sufficient for shadow mode
HistorianRetained process history, 6–12 monthsNeeded to build a credible baseline
Lab dataLIMS export or structured resultsRequired for Stabion and for validation
IdentitySAML or OIDC providerRBAC roles mapped during onboarding

AUDIT EXPORT

Everything is replayable

The audit export is the document your quality team will care about most. It contains every observation, prediction, proposal, confidence, citation, envelope, approver, model version and outcome for a batch.

It exports as JSONL for machines and as a rendered evidence pack for humans, and it is the same data either way.

  • JSONL export for SIEM and QMS ingestion.
  • Rendered evidence pack for inspectors.
  • Model and envelope versions pinned per event.
  • Tamper-evident chain across the batch.
audit export (excerpt)
{"t":"2026-07-14T09:12:04Z","batch":"B-24118",
 "type":"prediction","model":"emulson-wm-2026.07.03",
 "visc_eob":44900,"conf":0.91}
{"t":"2026-07-14T09:12:05Z","type":"proposal",
 "param":"homogeniser_rpm","delta":-40,
 "envelope":"env-v14","status":"pending"}
{"t":"2026-07-14T09:12:48Z","type":"approval",
 "actor":"a.rossi","role":"process_eng"}
{"t":"2026-07-14T09:12:49Z","type":"action",
 "written":true,"value":3160,"result":"ok"}

DOC PRINCIPLES

How we write documentation

VERSIONEDWITH THE PRODUCT
EXPORTABLEFOR VALIDATION
RUNNABLEEVERY EXAMPLE
HONESTABOUT LIMITS

DOCS FAQ

Using the documentation

ACCESS

Get the full documentation set

Evaluating Serumon seriously? We will provision documentation access for your engineering, quality and security teams under NDA.