Calibrate Your Policy

A confidence of 95% is only useful if answers given at 95% are right about 95% of the time on your data. This page shows how to check that for your own policy and pick an escalation threshold. You need a policy (Your first decision policy) and some labelled examples.

1. Label a sample

Collect real inputs with the answer you consider correct, one JSON object per line. Columns named expected_<question id> are labels; the rest are template variables:

{"id": "r01", "request": "Ordered a blue kettle, the box had a red toaster in it.", "expected_refund_eligible": "yes", "expected_reason": "wrong_item"}
{"id": "r04", "request": "Package never showed up, tracking stuck for 10 days.", "expected_refund_eligible": "no", "expected_reason": "late"}

Aim for 100+ labelled answers per question, including the hard cases you actually see; a sample of easy cases will look perfectly calibrated and tell you nothing.

2. Run the policy on it

Use the dataset runner from Run a dataset (set TEMPLATE_PATH, DATASET_PATH and your gateway URL). It writes results_receipt.jsonl with each question’s answer, confidence and hesitation. Keep samples at what you will use in production.

3. Read the report

python3 scripts/policy_calibration.py results_receipt.jsonl

Example from a 12-row refund dataset (illustrative size; recorded on a Vertex G4 endpoint):

== refund_eligible: n=12  accuracy=91.7%  Brier(top)=0.0769  ECE(10 bins)=0.0937
  confidence bin   n   mean conf  accuracy
       0.7-0.8    1      0.793     1.000
       0.9-1.0   11      0.993     0.909
  keep if hesitation <   kept   accuracy of kept   escalated
                    5%   75.0%             100.0%           3
                   16%   83.3%             100.0%           2
                   35%   91.7%              90.9%           1
                  any  100.0%              91.7%           0

• Accuracy: share of answers that match your labels.
• Brier (top answer) and ECE: how far confidence is from accuracy (0 is perfect). The reliability table shows it bin by bin: answers given at ~99% were right 91% of the time here.
• Hesitation thresholds: what happens if every answer above the cut-off goes to a person or a larger model. At 16% the policy handles 83% of answers on its own, all correct in this sample, and escalates 2 of 12.

The one error is instructive: “package never showed up” was answered eligible at 94% confidence, yet its hesitation was 34%. On a yes/no question 94% vs 6% is still a noticeable split, which is why hesitation (entropy scaled to the number of options) is a better gate than raw confidence. The label itself is also debatable: the policy text only mentions damaged or wrong items. Errors on a small sample are often policy wording, not the model.

4. Choose what to do with hesitant answers

Situation	Action
Few hesitant answers, errors are costly	Escalate above ~16% hesitation to a person
Many hesitant answers, a larger model is acceptable	Stage 2 cascade: "cascade_mode": "entropy" sends only hesitant items to Gemini (authoring guide)
Hesitation clusters on one question	Rewrite that question or its option descriptions; add missing options (taxonomy discovery)
You need an error bar per decision	samples: 4 adds standard error and agreement (~+35–40 ms)
Confidence is systematically too high or too low	Fit a temperature on held-out data (below)

What we measured on public benchmarks: an offline cascade escalating answers above 16% hesitation to Gemini reached 221/231 on JevBench with 39% escalated (Gemini on everything: 225) and 48/50 on the calibration suite with 34% escalated (EXP-14). Your escalation rate will depend on your data.

5. Temperature scaling (optional)

A temperature $T$ rescales probabilities after the fact: $T > 1$ softens overconfident answers, $T < 1$ sharpens them, and the answer itself never changes. Fit it on your own labelled sample, on data you did not use to choose it. There is no single served default: a temperature fitted on our other development suites and applied to a new one made calibration worse on 4 of 8 suites, and so did one per question type (PROP-20). Fitted on the suite itself (held out by 5-fold cross-validation), it helps on every suite.

Fit it. For the built-in suites, --auto-temperature fits $T$ by 5-fold cross-validation and scores each item with the $T$ fitted on the other folds (the same folds as the regression matrix). Its temperature_fit block lists the fold temperatures and raw vs held-out ECE:

./bin/dgem bench-calibration --from-receipt <receipt.json> --auto-temperature
./bin/dgem bench-jev --from-receipt <receipt.json> --auto-temperature

For your own policy, run it on your labelled rows (the dataset runner) and fit one $T$ per question type the same way. On our suites yes/no questions wanted about 3, choices about 1.3 and scores about 1.1.

Use it. Put the fitted value in the policy, or send it per request:

Where	How
Template	"temperature": 1.4 or "temperature": {"noul": 3.0, "choice": 1.3, "score": 1.1} next to "questions"
dgem decide	--temperature 1.4 (overrides the template)
Gateway POST /api/decide	"temperature": 1.4 or the per-type object (overrides the template)
MCP decide tools	temperature and temperature_by_type
Web Studio	Temperature field next to Prompt layout (also used by the batch runner)
/v1/systemone adapter	dgem systemone serve --temperature, or dgem serve --systemone-temperature

The temperature is applied after Stage 1 and after the Gemini cascade: the cascade still gates on raw entropy, and answers the cascade resolved are left as Gemini gave them. Responses report the temperature applied in temperature.

6. Re-check when things change

Re-run the report after changing the policy wording, the options, samples, or the serving image. A new vLLM version can change a few percent of individual answers even when overall accuracy is unchanged (image parity run).