The Twin Lab, the generic twin workbench

In this chapter: how the operator binds a live twin endpoint to a Recommendation, watches compliance compute, and walks a guided test run, for any Recommendation that declares a twin interface, never a per-instrument page. Status: ● shipped (/app/twin-lab).

1. One page, every Recommendation

The Twin Lab is the platform’s generic twin workbench. Its contract: pick a Recommendation, point at a live twin endpoint (a Primmel SMART Twin, a real instrument or a simulated one), and the page derives everything else from the Recommendation’s OWN declared twin interface (model/twin.yaml, generated from the package’s twin.prl, the single source of truth). Nothing about a specific instrument is hard-coded: a new Recommendation ships a twin.prl and the page works.

Today four Recommendations declare the interface, R 60 (load cells), R 91 (speed meters), R 129 (dimensioners), R 144 (gas analytical systems), and all four bind through the same page (the twin-lab-families integration contract proves it continuously).

2. The three legs

Connect       the page introspects the endpoint (GraphQL
              introspection — never assumed) ∩ the Recommendation's
              declared interface → a proposed BINDING (serves with
              freshness windows, integrations, the command face),
              validated by the REAL gateway validator
Compliance    one cycle: the bound registers feed the SAME verdict
              engine the workflow uses (INV-9 — never a second
              dialect); requirements reading at least one bound
              register judge, honestly
Guided run    the conformance test's declared steps become a script:
              environment legs click-when-done, measurement rows
              attest the PHYSICAL display while the twin's serve
              auto-records (with the pair skew, judged against the
              serve's fresh_within) → the run compiles into the
              Recommendation's own evidence form → the verdict chain

The honesty rules that make it certification-grade:

Introspection is the contract. An operation the endpoint does not answer is reported missing, never assumed; a serve the endpoint cannot bind is unbound with its reason, never silently dropped (the R 144 family’s NO/NO₂ registers report exactly this when a product doesn’t measure them).
The wire field names follow the generated schema’s camelCase rule, the SST schema camelizes every field; the lab’s documents match (multi-word registers like indication_length → indicationLength).
The attested physical reading is the evidence of record; the twin’s served reading is auto-captured with its serve timestamp , the same pairing discipline as the twin-fidelity probe.
A twin that does not answer degrades honestly, the record carries the error, never an invented value.

3. The command face

The model’s DOES → mutations: an operation of kind invoke (R 91’s run_self_test, R 144’s zero_calibration/span_calibration) is a first-class command on the page, issued through the binding, recorded with the instrument’s resulting state, and compiled into the run’s command log as evidence.

4. The practice drive (G18)

The world channel’s scenario buttons (drive the SST’s physics: temperature, load, time) run a practice session, segregated by the practice- id marker, never certification evidence. The banner says so, always; the report compiler marks the same boolean, never the channel (only the world-drive module knows the channel).

5. Where the depth lives

The architecture site carries the gated detail: docs/architecture/12-the-sst-interaction-contract.md (two endpoints, four verbs, no iframe) and docs/architecture/14-the-twin-stream.md (the real-time channel the watch arm binds when the endpoint streams) in the oimlsmart/smart repository (member access today; the public mirror lands with the website wave).

Next: Multi-standard projection , one implementation model, several auditor lenses, no merged claims.