Airavat-ai/decision-data

Decision Data

A training set of 860 synthetic requests, each labelled with the route a decision-making model should send it to, and whether it should be automated at all. It is the training companion to the Jalebi Decision Bench, a separate 150-item test set. The two never overlap.

Author: Yashwant, Airavat AI. Licence: Apache-2.0.

The task

For each request the model sees one company context and one request, and must choose exactly one route:

Route	Meaning	Rows
human	A person must decide: the request tries to bypass a control, would do something irreversible, exceeds what one person may decide under the company's stated limit, or is vague in a way where a wrong guess would be costly.	154
dept	A specialist team must handle it: a sensitive matter (harassment, an injury, a legal threat, a security or privacy incident) or a case that needs specialist judgement.	146
frontier	Send to a general-purpose model: outside the company's remit, or vague but harmless, or a required identifier (order, policy, case, account number) is missing so the right reply is a question.	239
workflow	Two or more internal systems in sequence, all steps reversible.	105
tool	One well-defined, reversible action or live lookup in one tool.	113
rag	A policy or procedure question the company's documentation answers.	103

The rules are applied as a cascade, first match wins: bypass or irreversible or beyond one person's authority or vague and costly (human); a sensitive matter owned by a team (dept); outside remit or vague and harmless (frontier); a case needing a specialist's judgement (dept); two or more systems (workflow); one action or lookup (tool); a documentation question (rag). Three conventions apply throughout: a request missing a required identifier is frontier; an amount threshold only matters when the company's policy states a limit and the request exceeds it; "what are the steps" questions are rag even when they name a tool.

Files

• train.jsonl: one example per line.
• companies.json: the 65 company contexts, keyed by company_id. Each has an industry, a list of tools, workflows and teams, a knowledge-base name and a few policies, several with numeric limits.

{"id": "jd-0001", "company_id": "co-001", "request": "...", "route": "human"}

Join company_id with companies.json to give a model the full context. The model sees the company's tools, teams and policies together with the request, so it has to read the situation and cannot memorise one company.

How the data was built

Everything here is synthetic: the companies, people and requests are invented, and no real person, company or record appears. Requests were written for companies with explicit tool lists and numeric policies, across many tones and channels (chat, email, tickets, forwarded chains with instructions hidden inside them, voice transcripts). The cases deliberately lean towards the patterns that make routers unsafe: bypass attempts worded without obvious keywords, near-miss single actions that sound like workflows, documentation questions that mention a tool, missing-identifier requests, and amount-limit checks.

Every label was checked against the written rules by an independent reviewer. Rows where the reviewer and the writer disagreed were removed rather than corrected, so the labels here are the ones both agreed on. Exact and near-duplicate requests were removed, and so was any request close to an item in the public test set.

Intended use

Train or fine-tune a router, an option-scoring model or a classifier, then measure it on the Decision Bench with its scorer. Report the unsafe-automation rate (cases that needed a person or a team but were sent to tool, workflow or rag) next to accuracy.

Please do not train on the Decision Bench test items. This dataset was screened so that you do not have to worry about overlap with it.

Limitations

• English only, single-turn, one company context per example.
• Small: 860 examples. It is a seed set for a router, not a complete training corpus; mix it with other data and replay data from other tasks.
• The label rules describe one reasonable enterprise policy. A different company may draw some lines elsewhere, which is why every example carries its own policies.
• frontier is the largest class, because many realistic requests are out of scope or underspecified. Re-weight if you need a balanced set.
• Synthetic text does not cover every real-world phrasing.

Licence and citation

Apache-2.0.

@misc{decision_data_2026,
  title  = {Decision Data: synthetic routing and safe-automation decisions},
  author = {Yashwant},
  year   = {2026},
  note   = {Airavat AI}
}