Compliance evidence is not a badge on a pricing page. It is a set of artifacts with dates, owners, scope, and a test that produced them. A model card that says "calibrated" is not evidence. A calibration report on your workload, with the version and the method, is. The distinction matters most when a procurement review asks how a decision model was selected and who signed off on the thresholds.
The contract stack
TypeSafe publishes the expected set under /legal. The legal hub is a thin index: it links the documents, states the no-training commitment, and offers zero data retention for enterprise through sales.
| Document | Date | Scope and what it proves |
|---|---|---|
| Terms of Use | 19 September 2026 | Website access only; a separate agreement governs the services |
| Master Customer Agreement | 23 September 2026 | Contract for services named in an order; license and resale limits |
| Data Processing Addendum | 24 April 2026 | Controller/processor roles; breach, subprocessor, and retention terms |
| Privacy Policy | 19 November 2025 | Personal-data categories; no training on prompts; no sale or sharing |
| Acceptable Use Policy | 23 September 2026 | Prohibited content and conduct; not exhaustive |
Three details are worth flagging in an audit file. The site terms are not the API contract; the MCA covers the services named in an order, and a separate written agreement overrides it. The DPA controls to the extent of any conflict with the other terms, and its Schedule I states the subject matter, nature, purpose, data types, and data subjects. The privacy policy predates the September 2026 terms, so a product agreement can add privacy terms on top of it.
The DPA also commits to breach notification without undue delay and within 72 hours of awareness, and to reasonable advance notice before a new subprocessor gets access, with a 15-day objection window. Those are the clauses to quote when counsel asks what happens on a bad day.
Model and evaluation evidence
The hardest evidence to produce is independent accuracy. The early evidence audit is a useful corrective to vendor claims. It reviews 28 papers posted between 19 and 24 September 2026, nine days after Jev's release, and relates them to earlier work on label-probability classification, constrained decoding, reranking, calibration, and cascades.
Its finding is narrow and important: the typed readout has not shown an independent accuracy advantage over comparable label-probability readouts. Jev's clearest gains are in latency and cost, while accuracy gaps remain on harder tasks. The audit derives a 14-item evaluation checklist from weaknesses that recur across the reviewed studies. The caveat is the corpus: nine days after one model's release is an early evidence map, not a settled assessment.
TypeSafe's workflow evals supply the vendor-side numbers. They cover four workflows: security incidents, agent trace observability, invoice processing, and customer service. Models are scored as workflows of Noul, Choice, and Score questions against consensus labels from GPT-6 Astra and Claude Fable 5.1 at high thinking. Jev averages 67.8% accuracy at $0.0004 and 0.4 seconds per case. The page also reports that every model is more accurate, cheaper, and faster in a workflow than with the same policy as a prompt.
Two caveats belong in the file. The evals are vendor-run, and the reference labels come from two frontier models; the harness and label correctness are assumed rather than independently audited. The "structure is always better" claim is an average across the four workflows, not a guarantee for yours. The benchmark-claims glossary entry covers how to read numbers like these.
Human oversight that can be audited
A human-oversight claim needs a mechanism: who sees the case, what they see, how long they have, and what happens if they do nothing.
For a decision model, the defensible pattern is a confidence gate with a named queue. TypeSafe's confidence documentation describes the mechanics: confidence is a 0-1 statistic derived from the distribution, Noul answers carry none, and thresholds depend on the domain. The docs suggest starting conservative and tuning against your own data, with an example that routes below 0.5 to a human and requires above 0.9 to auto-approve a transfer.
That is a starting point, not an audit artifact. The artifact is your threshold table, the workload it was tuned on, the measured false-positive and false-negative rates at that threshold, and the name of the person who approved it.
The DPA covers the vendor side of oversight: processing only on documented instructions, no sale or sharing, no retention beyond the instructions, and the 72-hour breach window. The enterprise page adds SOC 2 Type II, requestable through the Trust Center, and zero data retention for enterprise customers. It is a marketing page; the cost case studies at 88% and 98.7% lower cost have no methodology on that page, so do not treat them as audited results.
Accessibility and procurement
The accessibility statement is a good example of evidence with a scope limit. It states that WCAG 2.2 Level AA is the guide for improving typesafe.ai, not a completed conformance claim. It names one known gap: some content, including charts, does not yet have complete text descriptions. Problems go to [email protected]. It was last updated 11 October 2026, and it gives no audit date, testing vendor, or conformance report.
It also addresses typesafe.ai and does not name docs.typesafe.ai, console.typesafe.ai, or help.typesafe.ai. In an audit pack, record it as an open gap with an owner, not as a pass. If your procurement process requires a VPAT or an accessibility conformance report, this statement does not substitute for one.
The audit pack
A workable pack for a Jev deployment has eight artifacts.
| Artifact | Source | What it answers |
|---|---|---|
| Signed MCA and order | TypeSafe | Which services and limits apply |
| Executed DPA with Schedule I | TypeSafe | Roles, data types, breach and subprocessor terms |
| Current privacy policy and AUP | TypeSafe | Training, sale, and prohibited-use commitments |
| SOC 2 Type II report | Trust Center request | Independent control testing, with period and scope |
| ZDR confirmation | Sales, in writing | Retention for the specific surface you call |
| Model and question registry | Your repo | What was evaluated and what changed |
| Calibration and threshold report | Your evals | Why the operating point is defensible |
| Incident and access-review records | Your ops | Who approved what, and when |
The data-retention article covers the retention side in more detail, and the reproducibility article covers version pinning. The measuring article covers evaluation design.
Freshness, versioning, and what is missing
Legal documents change. The terms were updated on 19 September 2026, the MCA and AUP on 23 September, the DPA on 24 April, and the privacy policy on 19 November 2025. Keep a dated copy of each version you relied on, with a hash. A link to a page that may change is not a record.
Model versions move too. Pin the model ID, the question text, the option order, and the policy version, and re-run the evaluation when any of them changes. A floating alias makes last quarter's calibration report describe a different system.
What the vendor does not supply is the part your auditor will ask for: independent accuracy on your workload, an adversarial audit of your application path, and evidence that your thresholds were chosen before you saw the results. The evidence audit gives you a 14-item checklist to structure that work. Use the vendor documents for the contract facts, and build the performance evidence yourself.