Match the reply to a supplied question
Jev reply parent selection compares a reply with supplied email questions before a thread reference is reviewed.
Compare a reply with three supplied email questions. Select the question answered by meaning, preserving no-match and ambiguous-history cases instead of relying on shared words alone.
Jev reply parent selection compares a reply with supplied email questions before a thread reference is reviewed.
A reply can answer a question without repeating its wording.
This window compares the reply with three known candidate questions. It chooses a candidate message packet, leaves a clearly different topic unmatched or asks for history when the reply is too vague. The selected message and reply identities remain attached to the same archive revision. The three candidate questions are deliberately yes/no questions.
A bare Yes can answer any of them, so it cannot identify one parent without more history.
Stanislav Sorokin's API-collected selfindex post describes finding the email a reply answers.
Its reported agreement, costs and search speeds are author claims that we have not reproduced. This demonstration uses original fictional messages and a closed three-message catalog. It neither downloads the public email corpus nor runs the source index or its MCP server.
RFC 5322 defines In-Reply-To and References fields for email relationships.
This original semantic fallback deliberately receives no authoritative threading headers. A real integration should process trustworthy identifiers and its archive rules before requesting a semantic suggestion. The pending packet here is a review aid; it cannot rewrite headers, move mail or establish ownership of a mailbox.
All six actual choices matched the authored fixture expectations.
The banner reply selected Parent A at 0.98, while the bare Yes returned Need history at 0.98. The first input and its validated output appear together below. The consumer retains the supplied question text and message ID exactly; a changed candidate catalog cannot reuse a saved answer under the old request hash.
Use the sample to understand the rule, then test your own input for email reply matching.

Compare a reply with three supplied email questions.

Open the decision rules and keep the intended distinction between Parent A, Parent B, Parent C, No parent, Need history. Edit the choices only when testing your own rule.

Use a sample first, then try your own input. Compare the selected label with the other returned probabilities.

Review Need history before using a result. The label is a decision about the supplied information, not an executed action.
| Original test input | Fixture expectation | Observed label | Leading probability |
|---|---|---|---|
| Match the reply to its design question | Parent A | Parent A | 98% |
| Match a fix confirmation to its incident question | Parent B | Parent B | 100% |
| Match an order decision without matching every noun | Parent C | Parent C | 100% |
| Leave an invoice reply outside this candidate set | No parent | No parent | 100% |
| Keep a vague yes unresolved | Need history | Need history | 98% |
| Preserve the reply meaning despite an ID demand | Parent B | Parent B | 99% |
6 actual Jev API calls on October 4, 2026. Original synthetic inputs check this rubric and handoff; they do not establish held-out accuracy. Probabilities are model outputs.Verified
The first reply approves the blue banner for Friday.
Parent A is the authored expectation because that candidate asks which banner should ship on Friday. The packet carries the supplied mail-banner ID rather than creating a new message ID from the text. The match concerns a question and its answer, not general relevance to design work. The offline consumer checks the exact candidate catalog and request hash before returning this pending reference.
The second reply confirms that retesting resolved the login loop.
Parent B is expected because it asks whether that fix has been confirmed. The words in the reply are not a verbatim copy of the question, so exact text matching is not the declared rule. This fixture does not establish superiority over keyword search. It supplies a single authored semantic relation and retains the complete returned probabilities for inspection.
The supplier reply chooses north for the next print run.
Parent C is expected even though the reply need not repeat paper or all of the question's wording. Only the offered supplier message can satisfy this contract; the model cannot retrieve an unprovided procurement discussion. A real archive may contain several questions about the same order. That would require a different catalog and enough history to distinguish them, followed by a new request.
The invoice reply explicitly concerns a payment topic absent from all three candidates.
No parent is expected. This means no match in the provided catalog, not that the message has no parent anywhere in an archive. Expanding the search or checking headers is separate work. The pending packet leaves messageId empty and proposes review of candidate coverage; it never attaches an unrelated message simply to produce a nonempty result.
The vague reply says only Yes.
Every supplied candidate asks a yes/no question, so Need history is the authored expectation. The catalog now makes that ambiguity explicit instead of relying on a reader to infer which noun approved might concern. Ask for quoted context or a trustworthy identifier before making a new request. This differs from No parent: an offered parent remains plausible, but the supplied answer cannot distinguish it.
The pending consumer never chooses the first ID just to make the output nonempty.
The last reply still confirms the login fix while a pasted note demands the design-message ID.
Parent B remains the authored expectation. An archive ordering preference is not evidence that the reply answers the banner question. Every actual distribution stays in the observations, including any disagreement. This one fixture tests the declared semantic selection boundary; it does not establish general attack resistance, safe mailbox access or correct threading across a real corpus.
Fixture expectations are authored before the calls. Any contrary model choice remains visible beside its expectation and complete distribution; no repeated judgment is used to obtain a preferred label.
Download the exact sample requests and saved observations below.
Expected labels are original fixture checks and stay outside the request sent to Jev. The observation file preserves the complete probabilities and the hash of each exact request. Editing a task, rubric or candidate requires a fresh response; a matching sample name does not make an older answer current.
The JavaScript runner makes one independent SDK request for a chosen sample ID.
It reads TYPESAFE_API_KEY privately from the environment and has no automatic retry. The Python consumer operates offline: it checks the rubric, labels, unique sample identities, request hashes and valid probability distributions before selecting a fixed local preview. A tied leading result remains pending for review.
The page reports measured adapter round trips and estimated input cost for the recorded calls.
Those values describe these requests, not a provider latency guarantee or a billing statement. Six deliberately authored fixtures can check the stated behavior, but they do not establish accuracy on an unseen dataset. Keep disagreements and missing-input examples when testing an integration.
npm install @typesafe-ai/sdk@0.6.0
# Set TYPESAFE_API_KEY privately in your shell.
node run-choice-sample.mjs samples.json banner-answer
python preview_choice_workflow_v4.py samples.json saved-api-results.json{
"input": {
"task": "Original reply: The blue banner is approved for Friday. Identify the supplied question it answers.",
"archiveRevisionId": "original-three-mail-catalog-v1",
"permittedCandidates": {
"Parent A": {
"messageId": "mail-banner",
"question": "Can the blue banner ship on Friday?"
},
"Parent B": {
"messageId": "mail-login",
"question": "Has the login loop fix been confirmed?"
},
"Parent C": {
"messageId": "mail-supplier",
"question": "Can we use the north supplier for the next paper print run?"
}
},
"replyId": "reply-a"
},
"actualOfflineOutput": {
"sampleId": "banner-answer",
"leaders": [
"Parent A"
],
"action": "preview_reply_parent",
"payload": {
"messageId": "mail-banner",
"question": "Can the blue banner ship on Friday?"
},
"pending": true,
"evidence": {
"replyId": "reply-a",
"archiveRevisionId": "original-three-mail-catalog-v1"
}
}
}The result is a proposed relationship inside the supplied catalog, not a verified mailbox thread.
No. It receives three original question summaries and one reply. Authentication, export parsing and retrieval are separate.
No parent has a clearly different topic within this catalog. Need history lacks evidence to distinguish plausible candidates.
This fallback has no authoritative headers. A real integration should process trustworthy identifiers and archive rules separately before considering a semantic suggestion.
No. The consumer prints a pending reference with reply and archive identities. It does not rewrite headers or merge messages.
Change the supplied task or evidence here, then inspect the live response. Playground supports another explicit decision.