Guardrailed agents for laboratory work

The queue already holds work that should not start from a blank page. An accession sketched from a manifest. A coding check before a message reaches a state feed. A validation outline. A note on a quality-control flag. In the same week, someone asks whether the assistant can send the electronic laboratory report, submit a claim, or declare an outbreak while the supervisor is off the bench.

A limit nobody can point to is not a control. The laboratory still needs a named person on release, on transmission, and on any alert that leaves the building. A guardrail does not make the draft true. The assistant is decision support. It does not diagnose, and this site does not call it a certified medical device.

Those are the signs.

Salus describes itself as AI-native and not AI-in-charge. That sentence is the whole design. Intelligence in the product is Conrad, one governed persona with a least-privilege permission scope, supervised by a human. He is the voice on the marketing site and the voice in the application. He drafts, flags, extracts, and monitors. He is not a second scientist, and he is not a chatbot pasted beside an unchanged LIMS.

A guardrail, in this product, is a limit you can point to. Conrad does not release a result on his own. He does not transmit an electronic laboratory report on his own. He does not submit a claim on his own. He does not declare an outbreak on his own. The home page states the same boundary in shorter form: he never releases or transmits alone. The platform page generalizes it. An agent does not release a result, file a claim, or notify a jurisdiction alone. If a demo appears to skip that gate, the demo is wrong, not the policy.

The assistant is present on purpose. A floating orb on every screen is Conrad, with a proactive tip, an operational pulse, and a one-click action, so a technologist does not have to leave the bench to ask. Presence is not authority. The click that matters is the human approval that commits work, or the human rejection that does not.

The artificial-intelligence module lists the exhausting work. Accessioning from a manifest can be drafted for a person to approve. Operational reports and verifications can be drafted. Out-of-range results, presumptive positives, quality-control drift, and turnaround-time risk can be surfaced so attention goes there first. HL7 and FHIR messages and LOINC and SNOMED coding can be validated before anything reaches a state feed. Test patterns can be monitored for anomalies, and an alert can be drafted for an epidemiologist to review.

The workbench is the same idea pointed at documents. A validation or verification outline covers accuracy, precision, reportable range, reference interval, and analytic sensitivity as a skeleton. An SOP outline covers principle through limitations, plus a competency checklist for that method. Jurisdictional mapping starts a federal baseline and leaves the state overlay blank on purpose. A weekly synthesis of the regulatory log suggests actions. Every one of those outputs is a draft. The page says a person still completes the science, and that in the product every draft is stored, queued, and accepted or rejected by a second human.

Westgard pass and fail is computed by rules, not by a language model. That distinction matters on a survey. The model may explain a flag and draft a note. The rule engine decides the flag. Mixing those up is how a laboratory accidentally claims that a chatbot ran quality control. Salus does not claim that.

Each persona has an autonomy level. The module names Suggest, Acts with approval, and Autonomous in scope. Conrad is listed as Acts with approval: he drafts, and a human still signs results and transmits ELR. A human administrator can set a persona Active, Paused, or Disabled from the AI Personas screen. Disabling a persona is a human-only action. The kill switch stays in a person's hands.

Scope mirrors staff roles. A persona is not supposed to be able to do something its human counterpart could not do, and the scope can be granted or revoked laboratory by laboratory. Every status, autonomy, and scope change is written to an append-only control log that records who changed what, and when. The same audit habit the laboratory uses for a clinical act is applied to the agent. That log is how you answer 'who let it do that' without reconstructing a chat transcript.

Autonomous in scope is a level the configuration screen can express. It is not a description of Conrad releasing results. Nothing on the product page authorizes an unattended release, an unattended claim, or an unattended outbreak declaration. If a laboratory's policy forbids even a narrow autonomous act, the administrator sets the level and the log shows that choice. Policy lives with the laboratory. The software has to be able to obey it.

It does not make the draft correct. A suggested accession can still have the wrong submitter. A suggested code can still be the wrong LOINC for the method you validated. A suggested alert can still be a seasonal rise rather than an outbreak. Human review is the control, which is why the product refuses to skip it. Decision support only.

It does not diagnose. Watching a pattern and drafting an alert for an epidemiologist is not a patient diagnosis and not a declaration. The epidemiologist reviews the evidence, opens an investigation, and authorizes escalation. The journey on the home page ends on that human act.

It does not replace the laboratory director, the quality system CLIA requires, or the medical director's judgment under state law. It does not hold a CLIA certificate. It is not offered as a certified medical device. JIL Sovereign Technologies, Inc. operates this host. salusagentic.com is part of JIL Sovereign Technologies, Inc., a Delaware corporation. SalusAgentic is a trade name of that company. JIL Sovereign Holdings, LLC (WY) owns the SalusAgentic software. Global Hands, Inc. (TX) gave a grant to JIL Sovereign Holdings, LLC (WY) to build salusagentic.com under its public good initiative. The cost to build SalusAgentic was greater than what Global Hands paid. Global Hands may benefit from the software because it gave Holdings that money. JIL Sovereign Technologies, Inc. (DE) operates salusagentic.com. Holdings owns the rights, maintains the code, provides Level 1 support, and does not operate the code. Global Hands, Inc. does not operate this site. The software is the public-health laboratory product for qualified public health laboratories. The agent is a feature of that product. It is not a licensed practitioner.

It is also not a back door into some other company's chain or token system. Signed exchange between laboratories, where this site uses the word federation, means a partner verifies a package. It does not mean an agent is writing to a public ledger, and it does not mean a JIL network operates Salus. Those are different organizations and different products.