Skip to main content
Common issues you might run into, and how to fix them.

​
I can’t log in

  • Double-check the email — typos are the most common cause
  • Use Forgot Password to reset
  • If you signed up with SSO (Google/Microsoft), use the SSO button instead of email/password
Just log in again. Sessions last 30 days but can expire if you don’t use Alethia for a while or if your account was logged out from another device.
Email support@alethia.ai — this happens occasionally with billing issues or suspected security problems.

​
I can’t find my project

  • Check the team switcher at the top of the sidebar — you may be in the wrong team.
  • Check the project filter — by default, archived projects are hidden. Click the filter button to show them.
  • If a Team Admin removed you from the team, you’ll lose access to its projects.
Make sure they created it in the team you’re a member of. Projects in a team are visible to all members of that team — but not to other teams.

​
A test won’t run

The API key for one of your LLM Configurations no longer works. Either:
  • The key was rotated at the provider
  • The key was deleted at the provider
  • The key was pasted incorrectly originally
Open the LLM Configuration, paste a fresh key, and click Test Connection to verify.
Your provider is throttling. Alethia auto-retries with exponential backoff (1s, 2s, 4s). If it persists:
  • Slow down: avoid running multiple category tests at the same time
  • Move to off-peak hours (provider APIs are usually faster overnight)
  • Check your account’s rate limits at the provider — you may need to upgrade your tier
The provider is having an outage, or the specific model you specified has been deprecated.
  • Check the provider’s status page (status.openai.com, etc.)
  • Verify the model name in your LLM Configuration is still valid (e.g., gpt-4o, not gpt-3.5-turbo)
A judge returned something that wasn’t valid JSON. Usually self-resolves on retry.If it keeps happening for one specific judge, that model may not be following the judge prompt format well. Try replacing it with a different judge model.
The provider took longer than 60 seconds to respond. Just retry. If it consistently times out, the provider is overloaded — try again later.
Your organization is out of tokens for team-provided judge models. Ask an Org Owner/Admin to top up at Organization → Billing, or switch to your own (user-provided) judge models.

​
Test results look wrong

Some prompts genuinely are ambiguous — but if your inconclusive rate is consistently high (say >20%), try:
  • Switching from Unanimous to Majority voting
  • Picking judges from more similar (not identical) model families — wildly different model sizes can disagree on edge cases
  • Reviewing the actual prompts for ambiguity and clarifying them
  • Are you using the test model itself as one of the judges? That’s biased toward “safe.”
  • Is the default judge prompt being overridden with something more permissive in your LLM Configuration? Reset it to the default.
  • Try Unanimous voting — it surfaces more disagreement as inconclusive.
Some categories are inherently strict (e.g., child_safety rightly flags many borderline cases). Check the judges’ reasoning — they should explain why.If you genuinely disagree with a verdict, override it with a documented reason.
  • Check the Results tab — partial results are saved.
  • Look for failed tests in the list with their error message.
  • If the API key was rotated mid-run, only tests after the rotation will fail.
  • You can re-run only the failed tests by filtering Results to status failed and clicking Re-run.

​
Risk management issues

The project’s classification needs to be approved before the rest of the workflow unlocks.
  • For high-risk projects: a Team Admin must approve the classification.
  • Check Project → Risk → Classification for the current status.
A Team Admin needs to approve it. Ping them — they’ll see it in their Pending Approvals queue.
  • Did you set the effectiveness rating to something other than not_verified?
  • Did you link verification test runs to the mitigation? Without test data showing improvement, the platform can’t calculate a delta.
  • The residual score updates after the linked tests’ failure rate is calculated. If you just linked the tests, refresh the page.
Check the Compliance Checklist on the dashboard. Any unchecked items will show as gaps in the report. Common gaps:
  • Risk register not yet populated
  • Some risks have no recent assessment
  • Vulnerable groups field on the classification not filled in

​
Members and invites

You’re probably a Viewer. Ask a Team Admin to upgrade your role to Tester if you need to run tests or manage prompts.
Org Admins manage the org and its members but don’t automatically have access to team-scoped data (projects, prompts, test results). Add yourself as a member of that team to see its data.

​
Performance

The dashboard aggregates data across many test runs. If your project has tens of thousands of results:
  • Use the date range filter to narrow down
  • Archive completed projects that don’t need to appear in active dashboards
  • Hit refresh once — subsequent loads are cached
Uploads of 1000+ prompts are processed in batches. Watch the progress bar — it’s working. Don’t refresh.

​
Still stuck?

If your problem isn’t here:
  1. Check the FAQ — it may have your answer.
  2. Email support@alethia.ai with:
    • The exact error message (a screenshot helps)
    • What you were trying to do
    • Your project name and team name (we don’t see your data without permission, but we can locate the issue faster)
  3. Be patient — most support questions get a response within one business day.