Galileo, Cleanlab and Arize got acquired. Now what?

Six independent hallucination checkers were bought in fifteen months, most of them by platforms whose main product is something else. If one of them tells your team when an answer is wrong, here is what changes, what to ask your vendor, what to look for in an alternative, and what CalibratedAgents brings instead.

Who bought whom

In the last fifteen months, most of the independent tools that check AI answers for hallucinations were bought by bigger platforms. If your team relies on one of them to tell you when an answer is wrong, the checker now answers to someone else's roadmap.

This is normal for a young market: good tools get bought. But it changes what you are buying when you pick a hallucination detection tool today. Five of the six went into companies whose main product is something else: observability, security, data labelling, legal AI, a model lab. The checker is now a feature of that product.

Why it matters if you use one

A hallucination checker is a referee. It only works if it is independent and it shows up to every game. Acquisitions put both at risk.

• Roadmap risk. A product bought into an observability or security suite gets rebuilt around that suite. Standalone features, pricing and APIs can change or disappear. Humanloop's platform was shut down after its team joined Anthropic.
• Lock-in. The checker may soon work best, or only, inside the buyer's stack. If you do not run that stack, you carry the migration.
• Neutrality. When the company that checks your AI also sells the platform it runs on, or the model that wrote the answer, it is grading its own homework. A checker built from the producer's priors passes what the producer gets wrong; we measured this for LLM-as-a-judge, and the same logic applies to a checker owned by the platform.

Three questions to ask your current vendor this quarter: Will the standalone API stay, and for how long? Will pricing change when the integration lands? Can we export our labelled data and our thresholds if we leave?

What to look for in an alternative

Flagging a sentence that is not in the source is now table stakes. Cloud platforms do it too: Google Vertex AI's check-grounding scores each claim against a default threshold, and Azure's groundedness detection explains ungrounded segments for its Medical and Generic domains, optimised for English. The harder questions are these.

1. Is it calibrated to your domain? A score of 0.8 should mean the answer is right about 80% of the time on your data, not on a public benchmark. An off-the-shelf thermometer is fine; one calibrated to your furnace is better.
2. How many labels does it need? If adapting it takes thousands of labels or a fine-tuning project, most teams never do it, and the tool stays tuned to someone else's data.
3. Does it keep learning after you buy it? Your policies change, new document types arrive, products get renamed. A checker frozen at install drifts away from the thing it is checking.
4. Does it give a reason you can act on? The failing sentence, the source passage it should match, and why it does not, so a reviewer can confirm in seconds.
5. Does it check reasoning, not only facts? A step that does not follow, or a conclusion the evidence does not support.
6. Is it independent? No stake in which model, cloud or observability stack you run.
7. Does it fit regulated documents? Policies, contracts, financial and insurance documents, including non-English ones.

What CalibratedAgents brings to the table

CalibratedAgents is an independent verification platform for grounding and reasoning outputs, from models or from people. We built it around the seven questions above, so that the team reading the flags has less to re-read and less to worry about.

• Calibrated to your domain from about 200 labelled cases, with no fine-tuning, ever. It learns which signals matter for your definition of a wrong answer and sets its line to match. A flag at 0.9 is real about nine times in ten on your data, which is what lets you set a threshold and trust it.
• Learns in place as your domain moves. Reviewers' corrections feed back the same day: a new policy clause, a renamed product, a new form type. No labelled set to assemble, no retraining cycle, no frozen model between cycles. Why a fine-tuned checker cannot keep up.
• Pinpoint failures with reasons. The exact sentence or step, the source passage it should match, and why it fails: contradicted, unsupported, wrong value, wrong subject. On RAGTruth++, a check built to pinpoint the failing sentence found 59.6% of them at 75.5% precision, against 47.9% at 69.5% for an LLM judge calibrated on the same labels. Asked only whether the whole answer had an error, the two tied. The hallucination detection benchmark.
• Grounding and reasoning. Unsupported claims, broken logic steps and conclusions that do not follow from the evidence.
• Independent. Works with any model, any cloud and any eval or observability stack, and drops in as a scorer next to the tools you already run, including Vertex or Azure if you use them.
• In the flow, and over the archive. At most a second per check, so it sits where the output is made, and it re-checks what already passed, so drift in last month's answers shows up too.
• Built for regulated documents. Policies, contracts, claims, loan files and financial statements, including non-English sources.

What this means for the team: reviewers read what the check flags, with the reason attached, instead of re-reading everything. QC samples the riskiest answers first instead of a random five percent. And the record of what was checked, against what, is evidence you can show an auditor.

If your checker was just acquired

You do not have to decide this quarter. You do have to stop depending on a single roadmap you no longer influence. A four-week plan.

1. Week 1: export. Pull your labelled answers, your thresholds and your flag history out of the current tool while the API is unchanged.
2. Week 2: calibrate an alternative on 200 of those answers. Not a demo on a public benchmark; your answers, your sources, your definition of wrong.
3. Week 3: run both in parallel. Same traffic, both checkers, and compare at the sentence level: which flags were real, which errors each one missed, and how long each flag took a reviewer to confirm.
4. Week 4: decide with numbers. Precision at the recall you need, labels required, time to adapt to a change, and whether the vendor has any stake in the rest of your stack.

If your checker was just acquired, or you are choosing one now, send us 200 of your labelled answers. We will calibrate on them and show you, on your own data, what it catches and why.

Frequently asked questions

References

1. Dynatrace. Dynatrace to acquire Arize. Press release, August 2026.
2. Network World. Cisco to acquire Galileo for AI observability, April 2026.
3. Security Boulevard. Check Point acquires Deepchecks as it builds out agentic security platform, May 2026.
4. Cleanlab. Letter from the CEO: Handshake acquires Cleanlab, January 2026.
5. Artificial Lawyer. Harvey raises $550m at $15.5bn valuation, buys Guardrails AI, September 2026.
6. TechCrunch. Anthropic nabs Humanloop team as competition for enterprise AI talent heats up, 13 August 2025.
7. Google Cloud. Check grounding with RAG. Vertex AI documentation.
8. Microsoft. Groundedness detection. Azure AI Content Safety documentation.
9. CalibratedAgents. Hallucination detection benchmark 2026: RAGTruth++.
10. CalibratedAgents. Why outputs need independent verification.

Cite this

CalibratedAgents Research (2026). Galileo, Cleanlab and Arize got acquired. Now what? CalibratedAgents. https://calibratedagents.com/blog/galileo-cleanlab-arize-acquired-now-what