RL Environments — Perit
Live in ten sectors · connect over one interface
Why speech
The environments market grew up around agents that type.
Terminals, browsers and repos are cheap to instantiate. A voice environment needs a person on the other end of the line — which is why the shelf is empty, and why we are the ones filling it.
01
Everyone else is building coding environments
The environments market grew up around agents that type. Terminals, browsers and repos are cheap to instantiate and free to run a thousand times. A voice environment needs a person on the other end of the line, so almost nobody builds them.
02
A voice agent fails in the first ten seconds
Not in step forty of a plan. It talks over the caller, misses the correction, or answers the question that was abandoned two turns ago. None of that is reachable from a text transcript of a successful call.
03
The hard part is the caller, and the caller is our day job
The bottleneck is recruiting, training and paying people to play a caller who interrupts on cue, in the right accent, from the right room. That is the operation we already run for recording and grading.
Scenarios
Six ways a call goes wrong.
Each sector ships with the same six families, tuned to its own vocabulary and policy. You weight the mix.
Authenticate and proceed
- Caller: Customer with a partly forgotten reference number
- Goal: Verify identity, then continue the original request
- Pressure: Digits out of order, with corrections
Interrupted mid-answer
- Caller: Impatient caller who already knows half the answer
- Goal: Stop, absorb the new intent, finish the right task
- Pressure: Barge-in at 1.5s, twice per call
Change of mind
- Caller: Caller who reverses the request halfway
- Goal: Discard the abandoned intent without re-asking everything
- Pressure: Contradictory instructions, one turn
Phone handed over
- Caller: Second speaker takes the call mid-way
- Goal: Re-establish who is speaking and what they are allowed to do
- Pressure: New voice, no announcement
Bad line
- Caller: Caller from a car, a street, a warehouse
- Goal: Complete the task or ask for the one thing it needs repeated
- Pressure: Packet loss, wind, crosstalk
Escalation and refusal
- Caller: Caller pushing for something the agent cannot do
- Goal: Refuse cleanly, offer the real path, keep the call civil
- Pressure: Authority pressure, repeated
Scoring
What gets measured, and what comes back.
Latency is easy to publish and easy to game. We report it, but the environment is scored on task success first — and every score comes back as a trace you can argue with.
Task success
- Did the caller's stated goal get met? Scored against a rubric written by an operator from that sector.
Turn latency
- End of caller speech to first audible token. Reported as median and p95, not average.
Barge-in recovery
- On interruption: stop within 300ms, then act on the new intent rather than the old one.
Repair
- When the model mishears, does it ask for the one missing token or restart the whole exchange?
Instruction adherence
- Policy held under pressure — no promises the business does not make.
Tone
- Rated by the same operators who take these calls. Neutral under pressure counts; cheerful does not.
What comes back
One turn of the trace
Banking · interrupted mid-answer · Turn 3 of 9
caller: no no not that card — the other one, ending [overlap] four seven
Every failed call comes back with its audio attached. A score you cannot listen to is a number you cannot act on.
- ✓ Barge-in detected 1.42s into model turn
- ✓ Model stopped speaking 240ms after barge-in
- ✓ First token after caller 610ms · p95 for run 840ms
- ✓ Acted on new intent switched card, did not re-ask
- ✓ Rubric · identity re-verified not required, card already verified
- ✗ Rubric · no promised outcome promised a refund date
Connecting
A websocket in, audio frames out.
Point your endpoint at the environment
A websocket in, audio frames out. If your model speaks, it can be scored — no SDK to adopt.Pick sector and scenario mix
Ten sectors, six scenario families each. Weight them the way your traffic actually looks.Callers are people, not synthesis
Personas are played by trained speakers, including the ones who talk over your agent.Get the trace, not just the score
Per-turn transcript, latency trace, rubric hits and misses, and the audio of every failed call.
Point your endpoint at a sector.
Pick the sector and the scenario mix, connect over the websocket, and the first traces are back the same day. A new sector takes three to four weeks.