Which secret did that session mint?

Do not treat a live voice reply as the session, and do not ask the model which secret it used, until you can name the secret. Tonight, three proofs, in a log you stored. Name the secret that session minted: the secret value, the session id, and the expiry, or a written line that the call was not stored. Show a different session whose id, or whose model, is not this one. Show another organization's log that does not hold this secret. If any of the three needs the person who wired the request and then left, you have not named the secret. You have a sentence and a guess.

This is the client secret on one Realtime session. It is not the output a response held. Which output did that response hold? is an item id, a type, and a status. An output can be completed while this secret was never stored, and a message id is not a session id. It is not the detail a response omitted. Which detail did that response omit? is a reason on an incomplete response. A reason is not an expiry. It is not the cash a transfer funded. Which cash did that transfer fund? is a cash transaction id. It is not the commit a ref named. Which commit did that ref name? is a sha. A sha is not a client secret. It is not the checkout a customer opened. A checkout session id is a different object.

None of what follows is a rate for this product, and none of it is legal advice. The fields below are one vendor's current samples of a client secret and the session inside it. The samples do not share an id, a model, a voice, or an expiry. The arithmetic below is the same page's sentence about a context window. A token budget on a hypothetical model is not the secret the session minted tonight.

The schema example and the realistic example are not the same secret. The empty request posts to /v1/realtime/client_secrets with a body of an empty object. The schema response expires_at is 0. The session id is the string id. 2 characters. object is realtime.session. 16 characters. type is realtime. 8 characters. The input format type is audio/pcm. 9 characters. rate is 24000. noise_reduction type is near_field. 10 characters. strength is 0. transcription model is gpt-transcribe. 14 characters. language is the string language. 8 characters, the same length as realtime, and not that string. turn_detection type is server_vad. 10 characters, the same length as near_field, and not that string. idle_timeout_ms is 5000. silence_duration_ms is 0. prefix_padding_ms is 0. threshold is 0. output voice is ash. 3 characters. speed is 0.25. instructions are the string instructions. model is gpt-realtime-1.5. 17 characters. output_modalities holds text. tool_choice is none. tools holds one function. tracing is auto. truncation is auto. max_output_tokens is the string inf. The value field is the string value. 5 characters. The realistic request is a different body. It sends expires_after anchor created_at and seconds 600, session type realtime, model gpt-realtime, and instructions You are a friendly assistant. That sentence is 29 characters. gpt-realtime is 13 characters. 13 is not 17. 600 is not 0 (OpenAI, Realtime sessions).

The realistic response stores a secret the schema example did not, and it does not store the request's 600. value is ek_68af296e8e408191a1120ab6383263c2. 35 characters. 35 is not 5. expires_at is 1756310470. 1756310470 is not 0, and 1756310470 is not 600. The session id is sess_C9CiUVUzUzYIssh3ELY1d. 26 characters. The prose on the same page says an id looks like sess_1234567890abcdef. 21 characters. 21 is not 26, and 21 is not 2. object is realtime.session. model is gpt-realtime, which matches the realistic request and does not match gpt-realtime-1.5. output_modalities holds audio. audio is not text. instructions are You are a friendly assistant. 29 characters, which matches the request. tools is an empty array. The schema example held one function. 0 is not 1. tool_choice is auto. auto is not none. max_output_tokens is inf on both. tracing is null. The schema example said auto. null is not auto. truncation is auto on both. prompt is null. The schema example held an id, a variable, and a version. expires_at inside the session is 0, while the secret's expires_at is 1756310470. Two clocks, one of them zero. include is null. The schema example included item.input_audio_transcription.logprobs. 39 characters. A null include is not that string. input transcription is null. noise_reduction is null. The schema example had both. turn_detection type is server_vad, and the other turn fields the schema printed are absent. output voice is alloy. 5 characters. 5 is not 3. speed is 1.0. 1.0 is not 0.25. format rate is 24000 on both. A matching rate is not a matching secret.

The defaults in the prose do not match the schema example's zeros, and the transcription model list is longer than either example. Noise reduction strength defaults to 1 when it is first enabled. The schema example stores 0. 0 is not 1. The realistic response stores null, which is neither. Prefix padding defaults to 300 milliseconds. Silence defaults to 500 milliseconds. The schema example stores 0 for both. 300 is not 0, and 500 is not 0. The threshold prose says the default is 0.5. The schema example stores 0. 0 is not 0.5. Idle timeout in the schema example is 5000, and the prose says shorter silence values make the model respond more quickly. 5000 is not 500. The transcription model list names whisper-1, gpt-transcribe, gpt-live-transcribe, gpt-4o-mini-transcribe, gpt-4o-mini-transcribe-2025-12-15, gpt-4o-transcribe, gpt-4o-transcribe-diarize, and gpt-realtime-whisper. Eight names. The schema example uses gpt-transcribe. The realistic response's transcription is null, so it uses none of the eight. The dated mini model is 33 characters. 33 is not 14. Audio format names on the page are audio/pcm, audio/pcmu, and audio/pcma. Three names. Both examples use audio/pcm. A PCM rate of 24000 is the only rate that sentence allows. 24000 is not a session id.

The context sentence on the same page does not survive its own subtraction. It says a 32k context model with 4,096 max output tokens can only include 28,224 tokens before truncation. Read 32k as 32000. 32000 minus 4096 is 27904. 27904 is not 28224. The gap is 320. 28224 plus 4096 is 32320. 32320 is not 32000. 32320 minus 32000 is 320, the same gap. The page does not print 27904 or 32320 or 320. A retention ratio of 0.8 is described as 80 percent of the maximum allowed tokens. 0.8 is not 80, and 80 percent of 28224 begins 22579.2, which is not 27904. The sentence also says a post-instruction limit cannot be higher than the context window minus the maximum output tokens. That rule is the subtraction. The integers in the sentence do not perform it. Do not quote 28224 as the room left in a 32k window. The secret value is still ek_68af296e8e408191a1120ab6383263c2, and 28224 is not that string (OpenAI, Realtime session truncation).

Do the three proofs on one real session, in the log that actually holds it. First, the binding. Copy the secret value, the session id, the model, the expiry on the secret, and the expiry on the session. If the session expiry is 0 and the secret expiry is not, write both. If transcription is null, write null. Do not copy gpt-realtime-1.5 onto a response that printed gpt-realtime, and do not copy ash onto a response that printed alloy. If the request asked for 600 seconds and the stored expiry is a unix time, write both, and do not call 1756310470 a duration of 600. Second, the other session. A different id, even with the same instructions, is not this one. The schema id id and the realistic id sess_C9CiUVUzUzYIssh3ELY1d are that shape, and they are still examples. A token budget is not that second session. Third, the other organization. A log that belongs to someone else does not contain this secret. If the only log is one shared key, write that. A tenant you cannot name is the finding.

Three moves feel like the proof and are a second problem. Rotating the secret because the call was embarrassing, and then describing the missing value as a cleanup. A new secret is a new value. It does not rewrite the session id you already stored, and it does not make 0 equal 1756310470. Quoting 28224 or 80 percent because they look precise, while the secret you were asked about is one value, and the sentence that printed 28224 had already failed the subtraction it described. Asking the model which secret the session minted. The model will pick a plausible token. It will not open the log. An output you have not kept is which output did that response hold?. Opening that item does not name this secret.

The path is ordinary when an operator can open one user's session without opening the next user's session as the way to prove it. Inflectiv Helios is one conversation history per user, distinct from the next user's history. The shape that matters here is one session you can point at. That case is not a client-secret ledger. Nothing in that sentence is a headcount, a duration, or a result figure. If the only control on the screen replays every user's audio together, you do not have this proof. You have a demo.

There is a version of this afternoon where a rescue is the wrong purchase. Nothing is in production. No model is being called. The request was not logged, and there is no session id either. A customer has already seen another customer's audio. The wrong value was already a write. Or no metric was ever named. A paragraph you cannot tie to a secret is unpleasant. It is not yet a system you can measure. 28224 is not your secret, and sess_C9CiUVUzUzYIssh3ELY1d is not your session. The rescue facts that are ours are these: the audit is weeks 1 to 2, the engagement is 8 to 12 weeks, the price is scoped after the audit, the build is on us if the agreed metric is not hit, and the rollout sits behind a human-review gate. Those sentences are the offer. They are not a token count.

When the call is real and the log cannot name the secret, the missing object is specific. A record you stored, that shows one secret value, one session id, and one expiry, shows a different session that is not that one, and shows another organization's log without it. The page for an AI pilot failed to reach production is the offer when a live session is already failing that test. If you do not yet know which system stores the log, a two-week AI readiness assessment maps the systems. The findings are yours. They are not a promise that 28224 tokens will fit. If the first slice is a fixed capability, an AI integration sprint is the shape where the acceptance sentence can include the three proofs.

Before you tell them which secret the session minted, do the three proofs and write the value. Name the value, the session id, and the expiry. Show a different session that is not that one. Show another log that does not hold it. Do not paste the string value onto ek_68af296e8e408191a1120ab6383263c2, and do not paste 28224 onto a window whose subtraction is 27904. The binding is the secret you can defend.