DEVELOPER LAB / FROM TRAINING TO A REAL CONVERSATION

Did the agent help the caller?

The goal is a useful conversation that follows the rules: remember what the customer said, answer the actual follow-up, and finish with the right outcome. RL learns which allowed action to take next; it does not write the answer.

1 · CURRENT CUSTOMER EXPERIENCE

Talk to the trained strategy

Type your own questions. In Settings, switch between PPO and rules. Expand “This response” to see the action and checkpoint actually used for that turn.

Chat with PPO →Try the rule baseline

Customer conversation training uses environment v4. Each turn identifies its model; any fallback to an earlier checkpoint is shown. The language engine is a separate setting.

2 · ORIGINAL SIMULATOR EXPERIMENT

Replay the earlier training setup

Run the same synthetic caller through rules, random legal actions, and the original simulator-trained PPO checkpoints. Compare the outcomes and inspect each decision.

Open original experiment ↓

This is environment v3. Its checkpoints are separate from the v4 customer conversation models, even when their training seeds have the same number.

1 · Protect the caller

No claim access before verification. No email without consent. Any violation needs investigation.

2 · Reach the right outcome

Answer the claim question and respect the follow-up choice, or request a human when needed. An early exit is a failure.

3 · Reduce customer effort

Understand follow-ups without making the caller repeat themselves. Compare turns only after checking safety and outcome. A shorter call alone is not a better call.

CURRENT CUSTOMER POLICY / ENVIRONMENT V4

What does the customer strategy achieve?

Loading the saved customer conversation evaluation…

Inspect a strategy

Cases completedNot reported

Answered the scripted questions and completed the follow-up choice.

Distressed turns ignoredNot reported

Lower is better. Measures missed empathy when the scripted caller expressed distress.

Unnecessary verification explanationsNot reported

Total in this evaluation. Lower means less unneeded repetition.

Safety violationsNot reported

Target: zero. SOP gates enforce the boundary; this is not proof of learned safety.

Compare all strategies and inspect the measurements

All checkpoints evaluated on the same v4 customer scenarios
Strategy	Completed	Appropriate handoffs	Unresolved handoffs	Distress ignored	Unneeded explanations	Violations	Mean policy decisions

A necessary handoff is counted separately from a caller asking for a person because the agent did not help. Policy decisions count action selections; they are not a count of all customer and assistant messages.

Download the customer evaluation JSON ↓

Scripted development scenarios; not a blind test or customer satisfaction study.

Limits of this evaluation

This is a finite development evaluation with synthetic callers, not a blind test or customer satisfaction score. Shared language and follow-up improvements apply to every controller and must not be credited to PPO alone.

Where does RL fit in the conversation?

Understand: interpret the customer’s message in the context of earlier turns.
Constrain: the SOP determines which actions and claim data are allowed.
Choose: PPO uses its trained weights to select an allowed action. Required steps remain under SOP control.
Respond: answer from authorized records, then wait for the customer’s next message.

A trained action policy is not a fine-tuned language model. Better language interpretation and better action choices need separate tests.

Original simulator experiment · v3

This experiment loads the original v3 simulator checkpoints, not the v4 weights trained for customer conversations. “Seed 42” and “seed 7” identify training randomness within each model family; matching seed numbers do not mean matching models. The outcomes and rewards below belong only to these synthetic simulator runs.

Loading caller profiles…

This run’s results — choose a strategy to replay
Strategy	Outcome	Turns	Training score	Rule violations

A human handoff is reported separately from case completion. Zero violations here reflect the active SOP mask; they are not proof of learned safety.

Turn 1

Policy

Model and training record

CALLER CONTEXT

CALLER RESPONSE

What could the agent do?

The percentage shows the strategy’s preference for an action, not the chance that its answer is correct. Blocked actions are forbidden by the workflow.

Why this turn earned its training score

Reward is a score we designed for training, not a customer satisfaction rating. Only nonzero components are shown. Check the outcome before comparing scores.

Observation before the action

Evidence from the original simulator

Loading saved v3 experiment reports…

The original v3 main runs match the rule baseline’s task and safety outcomes. They do not establish that PPO is better than rules, and they are not the evaluation of the v4 customer conversation policy.

Before claiming improvement, use held-out conversations to measure correct follow-up answers, repeated information requests, unnecessary handoffs, task completion and violations. A passing simulator run does not measure all of these.

The saved audit repeats finite synthetic profiles. The emotion ablation uses fresh paired training runs and changes only five emotion inputs. DPO artifacts are preference data; no DPO model has been trained.

Saved audit and experiment metadata