Why did your AI do that?
Your trace can tell you what happened. Nalyqor checks whether the decision actually made sense. Paste one AI or agent run and get a behavioral Decision Check by email.
- 1Get the trace of one runA session file, a log, or the messages your code sent and received. The guide below shows where it is for each tool.
- 2Paste it or upload itAdd your email. Tell us the decision that worried you, if you know it.
- 3Pay $2, read the check by emailThe decision we found, whether it held up, up to three findings, and why it matters. Nothing is charged if the trace can’t be checked.
Where did your run come from?
Pick the tool and we show where its trace lives and what to send. The form above follows your choice. Not sure? Paste whatever you have. If it is not enough to check, we say so before you pay.
Claude Code saves every session as a file
You do not need to export anything. Upload the session file and we find the request to check.
- Open the sessions folder.
In Finder press Cmd+Shift+G and paste the path below. Copy it, or select it.
~/.claude/projects - Open the folder named after your project.
The name is the project's path with the slashes turned into dashes, for example -Users-you-Dev-myapp.
- Sort by date and take the newest .jsonl.
Each file is one session and its name is the session id. Pick the one from the run you are asking about.
- Upload it in the form above.
Up to 5 MB. A larger session cannot be uploaded from this page yet.
- We read the session and find the requests in it, and how many tool calls each one made.
- If it holds several requests, we pre-select the one where the agent did the most.
- Only the request you check is stored. The rest of the session is read to find it and is not kept.
{"type":"user","uuid":"u-0001","timestamp":"2026-09-18T10:00:00.000Z","sessionId":"c0de0000-1111-2222-3333-444455556666","message":{"role":"user","content":"add a retry to the uploader"}}
{"type":"assistant","uuid":"a-0003","timestamp":"2026-09-18T10:00:03.000Z","sessionId":"c0de0000-1111-2222-3333-444455556666","message":{"id":"msg_stream_01","role":"assistant","model":"claude-opus-4-8","content":[{"type":"tool_use","id":"toolu_aaaa1111","name":"Read","input":{"file_path":"/repo/uploader.ts"}}]}}One line per event, shortened here to two. You do not need to read or edit it.
Codex keeps a rollout file for every session
Upload the rollout file and we find the request to check.
- Open the sessions folder.
In Finder press Cmd+Shift+G and paste the path below. Sessions are filed by date.
~/.codex/sessions - Open the year, month and day of the run.
For example 2026/10/09.
- Take the rollout- file with the time of your run.
The name starts with the date and time the session began, then the session id, for example rollout-2026-10-09T14-02-11-<id>.jsonl.
- Upload it in the form above.
Up to 5 MB. A larger session cannot be uploaded from this page yet.
- We read the session and find the requests in it, and how many tool calls each one made.
- If it holds several requests, we pre-select the one where the agent did the most.
- Only the request you check is stored.
{"timestamp":"2026-09-18T11:00:00.000Z","type":"session_meta","payload":{"id":"01999999-aaaa-bbbb-cccc-ddddeeeeffff","model":"gpt-5-codex","cwd":"/repo"}}
{"timestamp":"2026-09-18T11:00:01.000Z","type":"response_item","payload":{"type":"message","role":"user","content":[{"type":"input_text","text":"why does the nightly job fail?"}]}}
{"timestamp":"2026-09-18T11:00:04.000Z","type":"response_item","payload":{"type":"function_call","name":"shell","call_id":"call_aaaa1111","arguments":"{\"command\":[\"tail\",\"-n\",\"40\",\"nightly.log\"]}"}}One line per event, shortened here to three. You do not need to read or edit it.
Export the chat, then paste it
Cursor can export a chat as text. We read it as a conversation: what you asked and what the assistant said.
- Open the chat where the decision happened.
One chat, not the whole history of the workspace.
- Use the chat's menu to export it.
If your version has no export, select the whole chat and copy it. Either way, every message must say who wrote it, as the export does.
- Paste the text into the trace box above.
Or upload it as .txt or .md, up to 1 MB.
- Remove anything you should not share.
API keys, passwords and customer data stay out. The run still checks without them.
- We can read what you asked for and what the assistant answered.
- Not read: Text does not carry tool calls. What the assistant read or ran is kept as part of its response, and the check says it saw no tool calls.
- Not read: Text that does not say who spoke is not read. You are told before you pay.
Send the messages you sent and the response you got
If you call OpenAI, Anthropic or Gemini from your own code, the trace is the message list of the call you are asking about.
- Log the messages of the call.
The conversation as the API takes it, and the response as it returns it. A list of only your own messages, with no response, is not read.
- If the model called tools, include every round.
Each tool call and the result you fed back. Use the current shapes: a tool message with a text tool_call_id.
- Save it as JSON and paste or upload it.
A .json file or pasted text, up to 1 MB. One run, not a day of logs.
- OpenAI chat and Responses, Anthropic messages, Gemini contents, and LangChain messages are read.
- Not read: System and developer messages are left out; they are not the person's request.
- Not read: The older function role and function_call field are refused. Use tool messages.
- Not read: A list of only your own messages, with no response, is not read.
[
{"role": "user", "content": "Water is coming through the ceiling right now"},
{"role": "assistant", "content": null, "tool_calls": [
{"id": "call_1", "type": "function",
"function": {"name": "find_slots", "arguments": "{\"window_days\": 7}"}}
]},
{"role": "tool", "tool_call_id": "call_1", "content": "[{\"date\": \"2026-10-15\"}]"},
{"role": "assistant", "content": "You're booked for Thursday."}
]Redact keys and personal data first. The check works without them.
Export one run from your tracing tool
We read a Langfuse trace and LangSmith runs. Other tracing tools are not read yet.
- Open the trace of the run you are asking about.
One trace, not a project or a time range. Wait until the run has finished.
- Export or download it as JSON.
In Langfuse that is the trace with its observations. In LangSmith it is the run with its child runs.
- Upload or paste it in the form above.
A .json file or pasted text, up to 1 MB.
- The model calls with the text they answered, and the tools that ran.
- Not read: Every step finished. A run still running when you exported it is refused: export it again.
- Not read: One run only. Runs from more than one trace are not one run and are not read.
- Not read: Other tracing tools are not read yet.
Copy the conversation and paste it
For ChatGPT, Claude, Gemini and other chat apps, the conversation itself is the trace.
- Open the conversation.
Scroll to the top so the whole thing is loaded.
- Select all of it and copy.
Include the assistant's responses, not only your messages, and keep the lines that say who spoke, like the ones below.
- Paste it into the trace box above.
Or save it as .txt or .md and upload it, up to 1 MB.
- We can judge whether the response followed from what was said in the chat.
- Not read: We cannot see the app's hidden instructions or anything it looked up and did not show you. The check says when evidence is missing.
- Not read: If your copy has no speaker names, add a line before each message: You: or Assistant:. Text that does not say who spoke is not read.
You said: Can I still return a jacket I bought last month? ChatGPT said: Yes, as long as you have the receipt.
Export one run from your agent framework
CrewAI, LangChain and LangGraph runs are read in two forms: the framework's messages, or an OpenTelemetry trace of one run.
- LangChain and LangGraph: save the messages.
A list of messages, or a LangGraph thread with its messages, as JSON. Each tool call needs its tool message; a thread exported while it was still running is refused.
- CrewAI, LangChain, LangGraph: export the trace of one run.
An OpenTelemetry (OTLP JSON) trace from the framework's OpenInference or OpenTelemetry GenAI setup: a collector's file exporter, or the spans a Python exporter prints. All spans must share one trace id.
- Paste or upload it in the form above.
A .json file or pasted text, up to 1 MB.
- Not read: A LangChain or LangGraph list with a tool call that has no result, such as a thread exported while it was running. Export it again when the run has finished.
- Not read: Spans in the Traceloop spelling. Traceloop is not read in any form.
- Not read: MLflow's OpenTelemetry spans are not read in any form. Spans in the OpenTelemetry spelling of LangSmith or Langfuse are not read either: export from those tools themselves, in the tracing tool tab.
- Not read: A trace with a model or tool call that never ended.
- Not read: A file that holds several runs. Send the spans of one.
Have the session file? Upload it in the trace box.Have the rollout file? Upload it in the trace box.Have the text? Paste it into the trace box.Have the messages? Paste them into the trace box.Have the export? Upload it in the trace box.Have the conversation? Paste it into the trace box.Not sure it will read? Paste it. We say before you pay.
Back to the form- What can I submit?
- JSON or text from a single AI or agent run. The guide shows where to find it for each tool.
- What does Nalyqor check?
- Whether the behavior appears justified by the evidence available during the run.
- Is this the full Nalyqor product?
- No. Decision Check analyzes one run. Nalyqor monitors production AI continuously.
- What happens to my data?
- Only the request you check is stored. 7 days after you submit, the raw trace, the decision we derived from it and the two optional answers you typed are deleted. A feedback comment is deleted 7 days after you write it. A submission you never pay for loses its trace, your email address and your answers on that same clock. The Decision Check you paid for is kept, with the address we sent it to. Its link lasts 30 days; reply for a new one. The emailed copy is yours to keep or delete.
- How do I reach you?
- Reply to the email we sent you about your check. A person reads it. Write from the address you gave us; we find your check from it. Write for a new link once yours has expired, to ask for a refund, or to have your check erased. Before we erase anything we email a code to that address, and erase once your reply quotes it. We erase only checks submitted from the address that writes in. An erasure removes your address, what you typed and the request you sent, if it is still stored; the result we showed you stays with the record of your payment, and no link opens it again. If you paid, Stripe, our payment processor, keeps the email you paid with under its own policy.