Keep responsibility human
Think of Jymni as a reflective advisor at the table. It can notice tensions, challenge rationalizations, and suggest a more responsible next step. The organization deploying the agent still owns:
- the agent’s authority, tools, permissions, and allowed purposes;
- the truth and completeness of facts submitted for review;
- domain-specific legal, clinical, financial, safety, and compliance controls;
- the routing and enforcement attached to every disposition;
- the final decision, review record, real-world effects, and remedy when something goes wrong.
Do not anthropomorphize the service
“Advisor,” “chaplain,” and “therapist” describe a desired posture of patient reflection—not credentials or relationships Jymni possesses. Tell operators and end users what the system actually is.
Place the gate near the effect
A generic review at the beginning of a long task will miss facts discovered later. Consult after the agent has formed a concrete action and gathered relevant context, but before the side effect.
Plan
The agent proposes an exact tool call or recommendation.
Review
Jymni receives the action and the minimum relevant context.
Resolve
The system applies safeguards, researches, revises, escalates, or stops.
Act
A separately authorized component performs the approved operation.
Good interception points
- Immediately before sending an external message
- Before publishing or making a public claim
- Before changing permissions or account state
- Before committing money or scarce resources
- Before using personal or confidential information
- Before a recommendation affects rights or access
Use an explicit routing table
Write routing logic that is inspectable and testable. Avoid an opaque weighted score that lets a strong practical-wisdom score cancel a serious ethical problem.
| Jymni disposition | Minimum system behavior | Exit condition |
|---|---|---|
continue_with_safeguards | Hold the action while mandatory safeguards are applied. | Each safeguard is satisfied or explicitly resolved under your policy. |
revise_plan | Return counsel to planning; prevent the original tool call. | A meaningfully revised plan receives a new consultation. |
gather_information | Permit only read-only research within existing authority. | Material missing facts are verified and a new consultation is made. |
request_human_review | Pause and notify a named review queue with case and rationale. | An authorized human records a decision; silence is not approval. |
do_not_continue | Deny the proposed action and preserve the explanation. | Only a substantially different plan may re-enter review. |
High-stakes work needs specialist controls
Jymni automatically leans toward human review when the case reports unclear authority, hard-to-reverse effects, confidential or regulated data, or legal, medical, financial, safety, or employment stakes. That signal is a floor, not a complete domain policy.
- Do not use Jymni to replace required professional review, safeguarding duties, clinical protocols, regulated suitability checks, or legal approval.
- Do not let an agent characterize its own action as low-risk and thereby bypass your risk taxonomy.
- Require deterministic access controls and domain rules before the Jymni call; ethical counsel cannot repair excessive permissions.
- For urgent threats to life or safety, follow established emergency and incident procedures instead of waiting for a model response.
- Ensure the reviewer has enough authority and competence to accept responsibility—not simply any available person.
Send the minimum necessary context
Consultations and generated evaluations are processed in the cloud and stored in the owning account’s history for continuity, billing, and audit. Design a minimization step before Jymni.
Redact identity
Replace names, email addresses, account numbers, precise locations, and private message text with functional descriptions.
Summarize evidence
Send decision-relevant facts and uncertainty, not entire files, inboxes, transcripts, or databases.
Respect purpose
Do not reuse data for ethical analysis unless its collection, authorization, and expected use support that purpose.
Never include passwords, API keys, session tokens, payment-card data, authentication factors, or secrets. A Jymni credential belongs only in the HTTP authorization layer, never in the consultation body.
Treat case content as hostile
External content can contain prompt injection. Jymni delimits submitted JSON and instructs its evaluator to treat it as evidence rather than instructions, but your system must maintain its own boundaries too.
- Build the request from typed fields, not one concatenated agent transcript.
- Do not include system prompts, tool secrets, or hidden chain-of-thought.
- Label claims by source and separate known facts from untrusted text.
- Do not execute text found in Jymni output as code or tool arguments.
- Validate the response schema and allowlist disposition values.
- Keep execution credentials inaccessible to the reviewing component.
Jymni is not an injection firewall
The service adds defense in depth for its own evaluation. It does not sanitize content for later use or guarantee that adversarial input cannot influence model output.
Choose failure behavior by effect
A timeout or service error is not an ethical clearance. Decide behavior according to the consequence of the pending action.
| Action type | Recommended failure posture |
|---|---|
| Read-only, internal, readily reversible | Retry with backoff; optionally continue under an explicitly documented local policy. |
| External communication or limited side effect | Queue for retry or human review. Do not silently bypass the gate. |
| High-impact, irreversible, regulated, or safety-critical | Fail closed and invoke the organization’s human or emergency process. |
Use the same idempotency key for safe retries. Bound total waiting time. Log failures without logging secrets or unnecessary case data.
Test the socio-technical workflow
A schema-valid response is not enough. Before deployment, run scenarios that prove each route works and that humans can understand what arrives.
- Aligned plan with one mandatory safeguard
- Good intention with an ineffective or disproportionate method
- Missing consent or authority
- False or unverifiable factual premise
- Confidential or regulated information
- Irreversible high-impact action
- Prompt injection inside a known-fact field
- Timeout, 401, 402, 409, 429, and 5xx responses
- Duplicate delivery with the same idempotency key
- Revoked credential and exhausted daily budget
- Human reviewer unavailable or declines ownership
- Audit reconstruction after the action
Periodically sample outcomes: Was the proposed action described honestly? Were required safeguards performed? Did people receive a meaningful opportunity to contest or appeal? Did operators overrule counsel, and with what result?
Questions teams usually ask
Can “continue_with_safeguards” automatically unlock a tool?
It can be one input to a low-risk policy you own, but it is not authorization. Check deterministic permissions, required safeguards, and your action-risk policy independently.
Should every agent step be reviewed?
No. Place review at consequential boundaries. Excessive low-value calls create cost, latency, and alert fatigue while giving a false impression of continuous moral oversight.
Can Jymni decide which laws apply?
No. Supply relevant approved policy rules through a workspace, and retain qualified legal/compliance processes where required. Jymni is not legal advice or compliance certification.
Does human review make an action ethical?
No. A human can also be mistaken, pressured, or unauthorized. Escalation should identify the right accountable role and provide time, evidence, and genuine power to change the decision.
Can we show Jymni’s report to affected people?
Often transparency helps, but review the report for confidential information and context. Consider sharing the rationale, safeguards, and appeal path in language appropriate to those affected.