← All tech signals
Infrastructure

Testing the Untestable

As models change weekly, checking whether an AI system still works is turning into a constant discipline rather than a one-time check before launch.

What is it?

Evaluation is turning from a launch gate into a running discipline, because models and the agents built on them keep changing after release. One track tests fixed capability; the other watches live behaviour, closer to monitoring than an exam. Whether either keeps pace with what it grades stays open.

Why now?

Weekly model churn broke the old assumption that testing happens once, before launch. Cognitive work is now cheap enough to run constantly, so checking it can run constantly too — a pre-release ritual becomes standing infrastructure. The question underneath: is this rigor, or just cost shifted onto every deployment forever?

Who’s building it?

Evaluation-infrastructure startups and funders, building the layer model providers and deployers now depend on.

BraintrustModel providersEnterprise deployersVenture fundersAgent builders
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — evaluation-infrastructure startups, funded to sit above the models.
Buyers — teams shipping AI products who need proof a change didn’t break something.
Gatekeepers — the investors and model providers who decide what counts as passing.
What’s at stake
Whether evaluation becomes its own durable category, or gets absorbed back into the model providers it grades.
Pricing power over what ‘working’ means for any deployed system.
Trust that regressions get caught before users notice.
What’s unsettled
Whether one model can reliably judge another’s output at scale.
How agent trajectories get scored when no fixed benchmark fits.
Whether continuous checking is rigor, or a cost nobody has priced yet.
02 · How the demand matureda progression, not a timeline
Necessity
Weekly model changes make a one-time launch test insufficient on its own.
Infrastructure
Funded platforms build scorers and judge-models directly into deployment pipelines.
Standard practice
Running checks on every release, not once before shipping, becomes the expected baseline.
Category
Enough funding and adoption to count as its own layer, distinct from the models graded.
03 · Where the market goesthe forks, not a prediction
→Standard or shortcut — whether LLM-as-judge scoring holds up as rigorous, or becomes a cheap stand-in nobody fully trusts.
→Capability testing and live monitoring stay two disciplines, or merge into one running check.
→Evaluation stays independent, or folds back into the model providers it exists to check.
→Agent evaluation matures on its own terms, or borrows benchmarks built for single-step tasks.
The open question is whether checking a system counts as knowing it works — or just as another layer of confidence nobody has actually tested.
04 · The AI Dev Process deskadjacent markets, moving now
Grading the Machine
Prompt testing turns into funded infrastructure.
Grading gets funded — an $80M Series B round.
SiliconANGLE
Keeping the Model Honest
LLM operations consolidates into fewer, bigger platforms.
An observability platform absorbed.
InfoWorld
Tracing the Reasoning
Observability extends from models to autonomous agents.
A platform launched to govern agents.
TechCrunch
05 · The Wire
01
A funding round for evaluation.
SiliconANGLE ↗
02
Investors bet on observability.
Axios ↗
03
Dynatrace agrees to buy AI evaluation firm Arize for $915 million
ITPro ↗
The panel on this signal
They agreed on
–Testing has shifted from a pre-launch gate to a permanent discipline tied to every deployment, not a launch event.
–Evaluation is becoming its own funded infrastructure layer.
–Agent evaluation genuinely needs new methods, not benchmarks built for single tasks.
The ML systems engineer disagreed on
–An LLM-as-judge score inherits the evaluator model’s own blind spots and calibration errors; without a human-labelled ground truth, a pass mark measures agreement between models, not correctness.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.