EnvironmentsFlagship
Test agents against a working copy of your systems.
- Real traffic, plus synthetic cases your data missed
- The agent runs the task end to end
- Failures point to the agent, or to the environment
The agent readiness platform
Know your agent is ready. Before it is your problem.
Stratix builds a working copy of the systems your agent touches, grades real tasks with verdicts nothing can fake, and puts every result on record.
Free to start. See it on your own tasks. No setup call.
Refund agent · a live check, in miniature
CRM
Refund policyrecorded
Orders
Order 4471 · $28recorded
Gift card · $12recorded
Payments
Charge · settledrecorded
Task 4471 · standardTask 4472 · harder
Agent
Verdict
Task 4471pass
Matched the known answer. On record
Task 4472fail
Agent missed it Claimed $40, policy says $28.
One standard task passes. A harder task fails, caught in the environment before it reaches a customer.
Sound familiar?
For every fear, a mechanism.
unverified
"The benchmark said 92%. Production disagreed."
verified
Environments, not benchmark quizzes.
Your agent is tested against a working copy of your actual systems.
unverified
"You found out from a customer, not a test."
verified
Before ship, not after impact.
Failures surface in a test that looks like your business.
unverified
"The judge agrees with itself, and nobody checks the judge."
verified
Verdicts are owned. Nothing grades itself.
Mechanical graders own every verdict. A leak gate blocks answer smuggling.
unverified
"The eval scripts live in one engineer's head."
verified
Records, not claims. Independent by rule.
Every result versioned, source labelled, on record. It outlives the script.
One platform
Three ways to prove your AI.
Environments, evaluations, and synthetic data. Use any one on its own. Combine them and every result gets sharper, and lands on the same record.
Evaluations
Grade any model or agent, on any benchmark.
- Your models and benchmarks, or auto-generated from your docs
- Graders, judges, and scorers, three ways to grade
- Compare models head to head, or across versions
Explore evaluations No environment required
Synthetic data
Generate test cases with known answer keys.
- Cover what your real data never captured
- No real or sensitive data required
- Feeds your evals and your environments
Ground truth included
Use one. Use all three. Each stands on its own. Together they close the loop from first test to signed record.
How it works
One environment. Five chapters.
Pick a service. Watch an agent run against a working copy of it, end to end. Click through it, or poke it.
Available environments
More on the way
Chapter 1 · BuildChapter 2 · RunChapter 3 · EvalChapter 4 · OptimizeChapter 5 · On record
A working copy of your systemsYour agent does real work in itEvery failure points somewhereReadiness climbs, on recordThe report a reviewer accepts
Your service's records assemble, every value labelled with its origin. Synthetic data fills the cases your real data missed, answer key included. A real task streams in. The agent reads field by field, acts, and a grader stamps the verdict against the known answer. One failure points to the agent. One points to the environment. Only an environment can tell you which, before a customer does. Same environment, same tasks, next version. You optimize, we prove the gain. The run chains into a readiness report, sealed on record. After launch, the same environment re-runs on every change.
Accounts
Account · Acme Corprecorded
Owner · Dana Parkrecorded
Environment couldn't answer
Opportunities
OPP-4471 · $28krecorded
Stage · Negotiationrecordedhidden
Contacts
Primary · J. Riverarecorded
Email · verifiedfilled in
Renewal case · missing from your data, filled in with the answer key
filled in
update OPP-4471update OPP-4471Tasks 4472 · 4473Same 40 tasks40 tasks · 3 models
Agent missed it
Sales ops agent
VerdictVerdictEval · 3 tasksReadiness report
Environment ready
The working copy is built and the answer key is minted from recorded values. Nothing has run yet, so there is no verdict. Send a task to grade it.
Task completed against the working copy. Matched the known answer.
pass On recordTask 4471 · matched the known answerpass
Task 4472 · the agent's answer did not matchfailpass
Task 4473 · a required field was not recordedfail
Points to the agent, not your environment. The twin served recorded values; the claim did not match.
Points to the environment, not your agent. The working copy refused to make something up. Record the field, the test gets stronger.
expected value · recorded · agent's answer · mismatch
required field · not recorded · no ground truth
Sales ops agent · Readiness report
Model A
88%
Model B
84%
Model C
71%
Ready under budget
On record
One agent, one environment, end to end. Left and right arrows change chapters.
1 · Build
CRM
Orders
Payments
Gift-card casefilled in
A working copy of your systems assembles. Every value labelled: recorded, or filled in with the answer key.
2 · Run
Task 4471 · refund $28
Refund agent
Refund posted · $28.00pass
The agent reads the twin field by field and acts. A grader stamps the verdict against the known answer.
3 · Eval
Task 4472fail
Agent missed it
Task 4473fail
Environment couldn't answer
Two failures, two causes. Every verdict points to the agent, or to the environment. That difference is everything.
4 · Optimize
v1 61%
v2 74%
v3 88%
Re-run the next version in the same environment. Readiness climbs 61 to 88. You optimize, we prove the gain.
5 · On record
Model A88%
Model B84%
Model C71%
Ready under budgetOn record
The report a reviewer accepts, sealed on record. After launch, the same environment re-runs on every change.
Evaluations
Grade any model or agent. No environment required.
Bring what you're shipping, then watch one answer get graded three ways that can't grade themselves.
$40≠$28The agent returned $40. The answer key says $28. Every grader below catches the gap.
One answer, graded three ways
Under test
Refund a cancelled insurance policy
Agent returned$40.00
Answer key$28.00
The same input, sent to each grader type on the right. Figures shown are illustrative.
expected $28.00returned $40.00
Graded 5 times · identical every run
FAILFAILFAILFAILFAIL
A verdict you can reproduce and defend. No model in the loop, nothing to drift or game.
expected $28.00returned $40.00
Same answer, 5 times · the verdict wobbles
FAILPASSFAILFAILPASS
Let the model check itself and the answer slips through half the time. Nothing to reproduce, nothing to trust.
The model never grades its own work. Pick any of the three and the verdict is owned, not self-reported.
Bring your own
Private & fine-tuned models
Benchmarks as CSV or JSON
Auto-generated from your docs
Compare & track
Compare models head to head
Re-run to compare versions
The difference on record
Where the failure happens is everything.
Same agent, same mistake. See where it surfaces.
Without Stratix
You ship the agent. Its flaw ships with it.
The errorRefunds $40 on a $28 policy
The cost$12 a refund, thousands a quarter
The trustA customer flags it, in production
An incident and a support queue. Unverified.
With Stratix
You test the agent. Its flaw never ships.
The errorFlagged $40 vs the $28 answer key
The cost$0 reaches an account, ever
The trustThe record flags it, before ship
A clean release, with the evidence attached.On record
Independence
Why the evidence holds up.
Mechanisms, not adjectives. The record does not perform.
01
Neutral by design
No vendor revenue share. No paid placement. Nobody pays to be evaluated.
02
Reproducible or it does not count
Every result can be rerun and returns the same answer.
03
Nothing grades itself
Mechanical graders own the verdict. A leak gate blocks smuggled answers.
04
On record
Published complete or not at all. Corrections logged, never silently edited.
We do not rank models. We verify agents in your environment.
About LayerLens
Verification infrastructure for AI.
LayerLens builds the independent proof layer that turns AI claims into evidence anyone can rely on. Stratix is our evaluation platform, independent by rule, with every result on record.
Neutral by design No vendor revenue share. No paid placement. Nobody pays to be evaluated.
2,000+Evaluations run to date
160+Models covered, independently
52+Benchmarks covered
On recordEvery result, reproducible
We evaluate the models your agents are built on
Trusted by teams building with AI
What your team gains
Built for the people on the hook.
From the blog
What we're learning, in the open.
Get new posts by email
Testing ends at your boundary. Evidence doesn't.
Prove your agent before it ships, then keep proving it in production. No claim without a record.
You pay for evaluation compute. Not for seats.
$2.10 spent
cap $5.00
40 tasks, 3 models, stopped under its cap.illustrative