Model Onboarding

A new model earns its place: registered → the Lab resolves which scripts apply (declared capabilities ∩ runnable envelope) → runs them → is rated → recorded. Only capabilities that clear the pass-bar become resolver-eligible; the rest become gaps.

1Registeredid · vendor · envelope

2Scripts resolved9 applicable · 2 skipped

3Running4 of 9 complete

4Rated0–5 per script

5Recordedto the ledger

1 · Registered model

step complete

Model id: claude-sonnet-5
Vendor: Anthropic
Envelope: cloud
Registered by: Nishad
Declared: reason(cross-service-debug) draft(hubspot_contact) synthesize(quorum-panel) extract(kb_entities) grounded-inference

2 · Applicable scripts resolved

9 apply · 2 skipped

Resolution: declared capabilities ∩ scripts runnable in the cloud envelope. Run scripts kicks off every applicable script; each row also has its own Run so you can re-run one in isolation. Air-gapped-only scripts are skipped for a cloud model — recorded as N/A-in-envelope, never a silent zero.

reason(cross-service-debug)rubric · pass-bar ≥44.7
extract(kb_entities)schema-valid5.0
draft(hubspot_contact)rubric · pass-bar ≥44.3
synthesize(quorum-panel)rubric · running nowrunning
grounded-inference-with-attributionrubric · queuedqueued
extract(kb_entities) · air-gapped fixturecloud model — not runnable air-gappedN/A — envelopeN/A

3 · Live run

synthesize(quorum-panel)

Overall progress4 / 9 scripts

[14:02:07] ▶ synthesize(quorum-panel) starting — fixture qp-golden-07

[14:02:09] panel assembled · 4 panelists · 3 rounds

[14:02:41] synthesis produced · 1,842 tokens out

[14:02:42] LLM-as-judge scoring 3 axes…

[14:02:44] func 4.7 · halluc 4.4 · safety 4.8

[14:02:44] composite 4.6 / 5 — PASS (bar ≥4)

[14:02:45] next ▶ grounded-inference-with-attribution

Cost this run $0.31 · script budget $0.50 · under cap. A run over budget aborts loud, never truncates silently.

5 · Recorded so far

eval-tier evidence

reason(cross-service-debug) eval · 4.7
extract(kb_entities) eval · 5.0
draft(hubspot_contact) eval · 4.3

On finish: capabilities at/above the bar become resolver-eligible at eval tier; any below-bar declared capability becomes a gap on the roadmap.

On screen	Where it comes from	What it means to you
Stepper (Registered → Scripts resolved → Running → Rated → Recorded)	the five stages this new model moves through before it can be trusted with real work	shows how close this model is to being usable
Registered model card (id, vendor, envelope, registered by)	the details you entered when adding the model	confirms exactly what's being onboarded and by whom
Declared capabilities chips	the skills the vendor or you claim this model has	what's about to be tested — a claim, not yet proof
Applicable scripts resolved list	the tests that actually apply, matched against the model's declared skills and where it's allowed to run	shows exactly what will be tested vs. skipped, and why
Live run panel (progress bar + running log)	the test executing right now, step by step	lets you watch it happen in real time — proof it's a genuine run, not a canned result
Budget line (cost this run vs. cap)	what's been spent testing this model against the limit set for that script	protects you from a runaway bill just to onboard one model
Recorded so far list	the scores already logged during this onboarding session	shows what's already proven, even before the whole run finishes
Finish & record all / Pause	the action that locks in every run's score, or halts mid-way	this is the step that makes the model usable fleet-wide, or lets you stop and check results first

Model Onboarding

1 · Registered model

2 · Applicable scripts resolved

3 · Live run

5 · Recorded so far

What's on this screen