Compare
The market is spending heavily on recording what happened. TruthLock is the layer that decides whether it should have, and proves the decision. Here is what each category does and does not do.
| Question | Observability | Public-web fact-check API | Groundedness guardrail | Lineage and signing | TruthLock |
|---|---|---|---|---|---|
| Records the prompt and response | Yes | No | No | Hashes only | Yes |
| Splits the response into claims | No | Yes | Sometimes | No | Yes |
| Grades against your own sources | No | No, public web only | Against retrieved context | No | Yes, ranked |
| Knows which source outranks which | No | Its own authority score | No | No | Declared hierarchy |
| Catches a faithful answer built on a superseded document | No | If the web knows | No | No | Yes, superseded_source |
| Says why a claim fell short | No | Score and prose | Score | No | Enumerated reason code |
| Refuses to round up on inexact match | No | No | No | No | Deterministic resolver |
| Decides what the claim may be used for | No | Returns a boolean | Block or pass | No | Per grade, per operation |
| Routes to a human with the record attached | No | Your code | No | No | Review queue, adjudication attested |
| Snapshots evidence at grading time | No | Citations | No | No | Hashed, in your bucket |
| Proves the record is unaltered | No | Shareable link | No | Yes | Merkle root, KMS signature |
| Proves the grade was permissible | No | No | No | No | Policy version in the record |
| Carries a lifecycle after signing | No | No | No | No | Supersession chain |
| Runs inside your boundary | Some | SaaS only | Some | Some | Dedicated AWS stack |
In particular
These are good products, and TruthLock borrows from them. They answer a different question.
Extract claims from any text. Research each against sources they find on the public web. Run several models against each other. Return a verdict, a score, a confidence, and the citations. Ship a CLI, SDKs, and workflow integrations. Some offer a financial warranty on the verdict.
If your question is "is this claim true according to the internet," use one.
Grade against your service manual, your policy set, your case file, or your canon. Several state plainly that they do not compare the answer with your own documents. Know that your bulletin outranks the forum thread. Tell a downstream system what it may do with a "mostly true" claim. Route anything to your reviewers. Sign the standing, or carry a record forward when the source changes. Run inside your account.
A score of seven out of ten describes an answer. It does not establish standing, and it cannot be defended as standing later.
In particular
Arize, LangSmith, Langfuse, Comet, Weights and Biases, and the APM vendors racing to catch up. OpenTelemetry instrumentation is table stakes, and TruthLock emits it. OWASP's Agent Observability Standard goes further and specifies hooks for hard controls on what an agent may do.
None of it declares which source wins when two conflict. None of it grades per claim. None of it distinguishes showing a claim with a caveat from acting on it. And the LangChain survey of more than 1,300 practitioners found 89 percent with observability already in place while quality remained the top barrier to shipping. That is not a tooling gap. It is a category error. A transcript is not an audit trail.