elan1 DECIDE · CALIBRATION LAB
A confidence score means nothing until it is tested on your cases.
Calibrated confidence means that when a model says it is 90% sure, it is right about 90% of the time on your own cases. The Calibration Lab in elan1 Decide tests and calibrates every decision for your tenant before any answer is automated. A decision without calibration is never routed to automatic.
No accuracy figure is quoted until it is measured on your cases. The chart beside is illustrative.
Illustrative specimen. Not customer data.
Stated 85% · observed 73% before, 84% after.
After a per-tenant fit, no bin sits more than 2 points off the diagonal; before calibration, 4 of 5 did.
Illustrative chart: calibrated confidence against outcomes. Before calibration the model is overconfident, right 80% of the time when it states 95%; after a per-tenant fit, stated confidence tracks observed accuracy.
Source: illustrative values drawn for this page. Accuracy is measured on your own cases.
| Stated confidence | Observed, before calibration | Observed, after a per-tenant fit |
|---|---|---|
| 55% | 60% | 56% |
| 65% | 64% | 66% |
| 75% | 68% | 74% |
| 85% | 73% | 84% |
| 95% | 80% | 94% |
01What it does
When the model says 90%, it should be right about 90% of the time.
The Lab builds golden sets of labelled cases with you. It tests candidate models against them, fits calibration per tenant and sets thresholds against a cost matrix you sign. Outcomes are joined back to decisions, so calibration is refitted as your cases change.
Who it's for
- Decision owners who must know how sure an answer is before it acts.
- Model-risk owners who need a dated record of how each model was tested.
03How it works
A model reaches your cases in seven steps.
Open a step to read it in place.
- Register it through the model adapter contract (MAC/1): typed request, typed answer, pinned version, region, licence, price and an evaluation card.
- Contract tests: does it return the declared shape?
- Offline bake-off against your golden set.
- Calibration fit, per tenant.
- Shadow: about two weeks beside your current process, with no effect on any case.
- Canary, per schema: 5%, then 25%, then 100% of cases.
- Promote. The previous model stays ready for 30 days.
When a calibration expires
When a calibration expires, the Router sends that schema's cases to review until it is refitted.
Where outcomes go
Review answers and outcomes are joined back to the decisions that caused them, so the fit follows your cases as they change.
04In the console
Calibration Lab in the elan1 Decide console.
A screen from the elan1 Decide console design, shown as an illustrative specimen. Not customer data.
Before a single case is automated, each engine is measured on your own double-labelled golden set, calibrated and tested against what you run today.
- Golden sets are frozen by hash and split 60/20/20.
- Agreement (kappa), calibration error (ECE) and accuracy are shown against their floors.
- Cut-over readiness lists each gate, through non-inferiority and shadow.
What it connects to
Writes go through elan1 Govern.
The Lab tells the Router which schemas are calibrated, and each Ledger entry names its calibration version. A threshold change is a policy change, so it is also declared in elan1 Govern, which re-checks it on every write. Writes go through elan1 Govern.
What it costs
Calibration Lab is included in every elan1 Decide plan. Decide charges for one thing it can count: decisions.
| What | Price · excl. GST | Includes |
|---|---|---|
| Growth — self-serve and developer teams | ₹4,80,000 per year | Typical volume 1,00,000 decisions a month |
| Mid-market | ₹36,00,000 per year | Typical volume 15,00,000 decisions a month |
| Enterprise — BFSI and insurers | ₹1,50,00,000 per year | Typical volume 1 crore decisions a month; SSO and SLA |
| Decisions beyond the plan allowance | ₹30 per 1,000 decisions | |
| A write intent Govern refuses | 40% of a governed action | Metered by elan1 Govern, not by Decide; the decision itself is counted once, as a decision |
All prices are exclusive of GST. Prices are current list prices; GST is added at the applicable rate on the invoice.
- First purchase: an LLM Decision Offload audit, or a claims or AML lighthouse.
All prices are exclusive of GST. GST is added at the applicable rate on the invoice.
Status
- Available in every elan1 Decide plan.
- Accuracy is measured on your own cases. We quote no figure before then.
- Governed actions in Mode B, through elan1 Govern: available.
- Each new schema runs in shadow until it reaches parity, before any route acts on its own.
Available now. See what’s next
What it does not do
- It does not quote accuracy before it is measured on your cases.
- It does not let a model skip shadow.
- It does not let a number act alone.
- It does not share one tenant's thresholds with another.
FAQ
Calibrated confidence means that when a model says it is 90% sure, it is right about 90% of the time on your own cases. A raw model score is not a probability until it has been tested against labelled outcomes in this way.
Against a golden set of your own labelled cases, sized with you for the schema and risk tier. The Lab compares stated confidence with real outcomes, fits the calibration, and refits as new outcomes flow back.
It is never routed to automatic. Its cases go to review, where a person decides, until the decision is calibrated for your tenant; if a calibration expires, the same applies until it is fixed.
It depends on the schema and the risk tier. We size the golden set with you during the lighthouse, and we do not quote a figure before we have measured one.
Outcomes flow back and the Lab refits. If a calibration expires, cases go to review until it is fixed.
A model enters through the model adapter contract and the same seven steps. Nothing above it calls a model directly.
Next steps
Related reading
One write path, four layers
Work and Decide propose. Govern and Run assure. Only Govern allows.
// elan1 DECIDE · CALIBRATION LAB
Bring one decision to a lighthouse.
Pick a high-volume decision. We run it in shadow beside your current process before any route goes live.