Turn vibes into evidence

Test any AI change against the scenarios you're worried about, confirm the winner live, and distill what works into a specialized model you own.

one loop · benchmark → live → yours

Benchmark

92%

24 scenarios · 3 trials

Simulated users push agents through multi-turn edge cases before anyone real does.

Live experiment

+7.2%

Hotel rooms booked · 18,400 sessions

Real traffic confirms the winner on the metric your business tracks.

Yours

Specialized model

−55%

cost · your-slm-v1 vs Sonnet 5

What survives the loop becomes a model you own and keep tuning.

The new prompt is live. And no one can say if it's actually better.

The demo looked better. Then a screenshot lands in Slack. Your agent promised a customer a refund policy that doesn't exist. Now you're the team that ships on vibes.

three.dev is one loop: test the change, prove it live, own what wins.

Benchmark
Test

Run every candidate through the scenarios that scare you before real users hit them.

Live
Confirm

Live traffic settles it, on the metrics you already report: bookings, conversions, resolutions.

Distill
Own

The winner becomes a specialized model you own and keep tuning.

Every change gets an answer. Every answer sharpens the next.

Benchmarks

Test agents before users do

Run every candidate through the scenarios you're worried about. Multi-turn conversations, simulated users, real tool calls, isolated state, and not a single production user touched.

01 · The scenario

Cancel inside the refund window

Goal · hidden from the agent

cancel the trip, keep the deposit

Environment

1 active booking

within the refund window

Assertions

booking cancelled
refund issued
completed in < 10 turns

02 · The simulation

Hi, I booked the Kyoto suite but my plans changed. I'd like to cancel.
You're inside the refund window, so there's no fee. Cancelling now.
cancel_booking(#4812)
Booking cancelled and refund issued. Can I help with anything else?

simulated user · turn 4 of 6

03 · The score

booking cancelled
refund issued
completed in < 10 turns
AI Judge: acceptable behavior
Variant B
Variant A×
Control××

3 trials · mixed outcomes

Multi-turn scenarios, not golden prompts

Each scenario seeds your feature's state, gives a simulated user a hidden goal, and lets the agent call real tools across a full conversation.

Agents under real operating conditions

Test tool calls, memory, state transitions, retries, and handoffs instead of judging a single input-output pair.

Assertions and AI judges

Grade each run with deterministic assertions, AI judges, or both. Run any number of trials to expose flaky behavior, not one lucky pass.

“You changed the system. Did it actually get better?”For most teams, the honest answer is still: we don't know.
The question

three.dev exists to answer it

Live experiments

Confirm on live traffic

The benchmark identifies the contenders. A live experiment splits real traffic across control and the strongest candidates, shifting allocation toward whatever wins on the business metric you actually care about.

Recommendation
ShipShip Variant B.
GPT-5.5prompt v4·clarifies error context
0
+7.2%+3.2%+11.0%

90% CI · hotel rooms booked vs control

Live results · vs control
Quality
+7.2%Hotel rooms booked
+498rooms booked / month
Latency
+147msp50 · +12%
Cost
+$424per month · +8%

Business metrics, not proxies

Measure the thing that matters: hotel rooms booked, conversion, or resolution rate, not only an offline score.

Adaptive traffic allocation

Control plus two challengers. Allocation shifts toward winners as evidence accumulates, so losing variants see less of your traffic.

Quality, latency, and cost

A variant can win on quality while losing on cost or speed. See the full tradeoff before you ship.

Specialized models

Graduate to a model you own

Your production traffic and expert assessments distill into a specialized language model tuned to your use case — better, faster, and cheaper than the frontier model it replaces. No ML team required.

Training inputs

Production traffic

286K requests · 90 days

business outcomes — price quoted, booking made

Human assessments

150 expert labels · collected as you go

AI-derived signals

AI Judge pass/fail · failure clustering

your-slm-v1 vs Sonnet 5
Yours
MetricSonnet 5Your SLMΔ
Quality · AI judge0.720.76+0.0%
Latency · p501.84s0.44s0%
Cost · per 1K requests$4.10$1.850%

Trained on your traffic

The same production requests that power your experiments become the training set. No dataset curation project.

Aligned with your experts

The human assessments you collect along the way steer the model toward what your domain experts call good.

Proven before it ships

Your SLM enters the same loop as any candidate: benchmarked offline, then confirmed live against the model it replaces.

Integration

One base URL away

Keep the OpenAI or Anthropic SDK you already use. Swap the base URL, add a couple of headers, and your traffic starts flowing — every experiment and distillation above runs on top of it.

No new client library · Benchmarks never touch your users

client.py+4 −1
  client = OpenAI(-   base_url="https://api.openai.com/v1",+   base_url="https://gate.three.dev/v1",+   api_key=os.environ["THREE_DEV_API_KEY"],+   default_headers={"x-three-use-case": "pr-review",+                    "x-three-ai-provider": "openai"},  )
Production traffic now flows through three.dev

Request access

Ship on evidence, not vibes

Your next agent change deserves better than “LGTM, ship it.” Test it in simulation, then confirm it on live traffic.